The Atlas of the Agentic Economy

v0.3 — A practitioner's map of the labor market that didn't have a name yesterday

By Thomas Oxlee Founder, shipstacked.com Currently embedded as the AI integration operator at a regulated EU business under AI Act exposure

shipstacked.com/atlas

Foreword

CVs were invented in the 15th century. The job title as we know it — a stable, searchable, salary-banded category — is a 20th-century artifact, built for an industrial economy where roles changed on a timescale of decades. Most of the infrastructure we use to find work and find workers — LinkedIn taxonomies, ATS keyword filters, ISCO codes, recruiter Boolean strings, the standard CV format — assumes that what you did last year predicts what you can do this year, and that the role you're hired into has been done before.

That assumption broke in the last 12 to 18 months. It didn't bend. It broke.

We are now in a labor market where the most valuable people are doing work that did not exist as a recognizable role two years ago. The companies that need them often cannot describe what they need; they describe symptoms. Our Claude integration is bleeding budget. Our agent broke and we don't know why. Our automation half-works and is making a mess of our daily operations. We hired a consultant for €50,000 and now we have a sophisticated system that nobody can operate. The people who can solve these problems exist. They are not on LinkedIn under the right title. Many of them are not on LinkedIn at all.

On May 4, 2026, Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs announced a $1.5 billion venture whose stated purpose is to solve "one of the most significant bottlenecks to enterprise AI adoption — the scarcity of engineers who can implement frontier AI systems at speed." OpenAI announced a near-identical $4 billion venture with TPG and Bain the same day. On May 11, 2026, OpenAI formally launched The OpenAI Deployment Company and simultaneously acquired Tomoro — a 150-person UK consultancy — to staff it, because the labor cannot be hired through normal channels at the scale required. ManpowerGroup's 2026 survey of 39,000 employers across 41 countries found that AI Model & Application Development is now the single hardest-to-fill skill in the world, for the first time in the survey's history. Bain estimates half of 1.3 million US AI jobs may go unfilled by 2027. Forrester says 75% of organizations attempting to build AI agents in-house will fail. RAND data shows 80% of AI projects fail to deliver business value. MIT reports 95% of GenAI pilots never reach production.

The labor market has not produced a coordinated response to this because it cannot see the work clearly enough to coordinate. The roles are not named. The supply is not classified. The demand is described in symptoms.

This Atlas is an attempt to name what's happening. It is written by a practitioner — someone currently doing this work at an operating company, an EU regulated business navigating its first agentic deployment under AI Act exposure — not by an analyst observing it from the outside. It is intentionally early. It will be wrong in places. The point is not to be the final word. The point is to put down a stake while the labor market is still legible enough to map, and to invite practitioners to push back.

The Atlas comes in six parts because the labor surface of the agentic economy has at least six structurally distinct populations, and forcing them into one taxonomy obscures what's actually happening on the ground.

Part I — The Workforce. Twenty-eight specialist roles, organized into five clusters, that describe employed labor inside companies adopting AI. Each role gets an automation trajectory: how much of this work will collapse to one-human-plus-agents in the next eighteen months, and how much remains irreducibly human.

Part II — The Operators. A new economic unit that does not fit into employment at all. The solo or small-team practitioner who runs a portfolio of agents and rents the output to customers. Five operator types, each described as a business shape, not a job title.

Part III — The Compliance Layer. The roles emerging at the intersection of AI deployment and regulatory exposure. Restructured into three sub-clusters reflecting how frontier labs actually organize this work: Research-flavored, Operations-flavored, External-partnership-flavored.

Part IV — Alignment & Interpretability Research. A separate research-flavored population, connected to the Anthropic Fellows / MATS / Redwood / ARC pipelines, distinct from the compliance work in Part III. Different career path, different supply pool, different demand pool.

Part V — Model Training & RLHF. The Mercor / Scale AI / Surge AI population. Three tiers. Already a $1B+ market dominated by Mercor; described here for completeness, as it is part of the agentic-economy labor surface even where shipstacked does not directly compete.

Part VI — Industry Vertical AI Specialists. Domain practitioners who learned AI, rather than AI practitioners who learned a domain. Healthcare, legal, financial services, defense, manufacturing. Distinct supply pool, distinct compensation curves, increasingly the largest segment of AI hiring globally.

A few principles that shape what follows:

• Roles are described by what people do, not what they're called. If you do this work, this Atlas is about you.
• Demand and supply signals are observable. Where the demand surfaces and where the supply lives.
• The categories are practitioner-defined, not analyst-defined. Boundaries are blurry on purpose, because the work is blurry.
• Automation trajectory is honest. Several roles described as full-time positions today will be one human plus an agent fleet in eighteen months. Said so by role.
• The frontier labs are the leading indicator. Most of the structural specialization in this Atlas was first observed in actual hiring at Anthropic, OpenAI, Palantir, Cursor, Databricks, Cohere. What's at the labs now is at customers in 12-24 months.
• This is v0.3. Pull requests welcome.

What follows is the full taxonomy.

The five clusters

Cluster A — Implementation & Deployment. The Applied AI super-cluster. Forward-deployed work, integration work, deployment work, transferable delivery work, deployment strategy, partner channel work. Most successful deployments at frontier labs use paired roles, not solo specialists. Headcount stable or growing across most of the cluster.

Cluster B — Reliability & Operations. The people who keep AI systems running once they're deployed. Substantial automation collapse expected in this cluster within 18 months — the work itself is increasingly agent-amenable.

Cluster C — Governance, Risk & Compliance. Summary here; expanded fully in Part III. The buyer is increasingly distinct from the rest of the Workforce, which is why the cluster has its own part.

Cluster D — Design & Architecture. The people who decide what AI systems should do. Includes the eval-design specialism that frontier labs treat as central infrastructure but that customers are only beginning to recognize.

Cluster E — Translation & Enablement. The bridge between technical capability and organizational reality. Implementation leadership, training, fractional executive coverage.

Notation

🔴 Resistant — Irreducibly human. Agents support but cannot replace the core function in the foreseeable horizon. Headcount stable or growing.

🟡 Partial — Team size collapses. What was a 3-5 person function becomes 1-2 humans plus an agent fleet within 18 months. Headcount contracts; core role survives.

🟢 Collapsible — Within 12-18 months, this is one human supervising agents. The role survives but the headcount per company drops dramatically.

Specialization axes

Three axes apply across most of Cluster A and parts of D and E. v0.3 makes them explicit:

Axis 1: Pre-sales vs. Post-sales. Anthropic's Applied AI Architect (pre-sales technical advisor) is structurally different from Anthropic's Solutions Architect (post-sales delivery). Different supply pools, different career paths, different compensation curves.

Axis 2: Customer Segment. Enterprise / Federal-Government / Mid-Market / Startups / Industries-Vertical / Partners-Channel. The work differs materially by segment.

Axis 3: Engagement Model. Advisory / Forward-Deployed / Paired-with-Deployment-Strategist / Partner-Channel / Embedded-Long-Term.

These are not separate roles. They are dimensions along which the same role specializes. The Atlas notes them per role where they're load-bearing.

CLUSTER A — Implementation & Deployment

A1. AI Integration Operator 🟡

What they do. Take an AI capability and integrate it into the operating reality of a specific company. Wire AI into existing workflows, existing tools, existing data, existing human processes. The work is roughly 30% engineering, 30% operations, 40% organizational translation. Not the same as a software engineer; not the same as a forward-deployed engineer at a frontier lab; closer to a fractional CTO with AI specialism.

What good looks like. Sits with a department head, understands the existing workflow in ninety minutes, identifies where AI provides leverage and where it introduces unacceptable risk, ships a working integration in days, not quarters. Obsessive about understanding the business before touching the technology. Knows that 80% of AI integration failures are workflow problems, not model problems. Carries a portfolio of integrations described specifically — what shipped, what broke, how it was fixed.

Demand signals. Mid-market companies that have bought AI tooling but cannot operationalize it. Job posts titled "AI Implementation Specialist," "AI Deployment Engineer," "AI Solutions Engineer" with the actual work being integration. PE portfolio companies under operating-partner pressure to deploy AI quickly. Any company that has done a 6-month AI pilot and now wants someone to "just make it work."

Supply signals. Engineers who have shipped multiple AI integrations across different SaaS products and can describe each one specifically. People who write public technical posts about connecting things — Stripe + Claude, Supabase + Anthropic SDK, n8n + custom MCP servers. Solo builders who have shipped a working AI-augmented internal tool at a non-tech company. Polyglot — comfortable across at least three or four different tech stacks.

Common failure when the wrong person is hired. Companies hire a "Senior Software Engineer" or "Solutions Architect" and end up with someone who can build clean abstractions but cannot navigate organizational politics, won't go talk to the operations manager, and ships a technically excellent system that nobody uses. Or hire a consultant who delivers a slide deck and leaves before the integration is operational.

Automation trajectory. Partial collapse. A 3-5 person integration team becomes 1-2 humans plus an agent fleet within 18 months — the agent fleet handles connector building, configuration, log analysis, and routine integration testing. The human work that remains: organizational translation, executive judgment, navigating workflow politics.

Specialization. Most commonly post-sales, mid-market, embedded-long-term. Less commonly pre-sales (where A2 or A6 fits better).

Adjacent roles. A2 (Forward Deployed Engineer, more senior, more technical). A5 (Agent System Integrator, transferable-delivery focus). A6 (Deployment Strategist, the strategy-flavored counterpart).

A2. Forward Deployed Engineer (FDE — AI flavor) 🔴

What they do. Embed inside a customer company — physically or via deep, sustained engagement — and build production AI systems alongside the customer's team. The role originated at Palantir in the early 2010s, where they were called "Deltas" and where, until 2016, Palantir had more FDEs than software engineers. Revived at scale in 2025-2026 as the dominant pattern for enterprise AI deployment. Indeed reports 800-1000% growth in postings between January and September 2025.

What good looks like. Bilingual: deeply technical AND able to operate in a customer's executive room without translation. Reads the customer's quarterly reports before the kickoff meeting. Makes architectural decisions a less senior engineer would defer to product. Ships into production within weeks, not months. Handles pre-sales scoping, post-sales implementation, integration, evaluation, monitoring, and ongoing iteration — sometimes simultaneously. From Palantir's own description: "FDEs responsibilities look similar to those of a startup CTO: you'll work in small teams and own end-to-end execution of high-stakes projects."

Demand signals. Direct: Anthropic ("Forward Deployed Engineer, Applied AI"), OpenAI (Forward Deployed Software Engineer plus vertical specializations: Life Sciences FDE, Semiconductor FDE, Government FDE), Palantir (the original FDSE role), Salesforce (committed to hiring 1,000 FDEs), Databricks ("AI Engineers, FDE"), Cohere, Ramp, Rippling, Intercom. The Anthropic/Blackstone $1.5B venture is fundamentally an FDE-deployment vehicle. The OpenAI Deployment Company (launched May 11 2026, $10B) acquired Tomoro to gain 150 FDEs from day one. EY launched a UK & Ireland FDE practice in April 2026. Indirect: mid-market companies in healthcare, financial services, legal, manufacturing, retail with a board mandate to deploy AI.

Supply signals. Ex-Palantir is the gold standard. Ex-frontier-lab applied AI teams. Senior engineers from AI-native startups (Cursor, Anysphere, Anthropic, OpenAI, Linear, Vercel) open to embedded work. Strong technical writing as a signal — FDEs have to communicate constantly with non-technical executives. People with public case studies of "I went into Company X and shipped Y in eight weeks."

Common failure when the wrong person is hired. Companies hire a senior backend engineer who is brilliant but cannot read an exec room. The customer relationship deteriorates within a quarter. Or they hire a former management consultant who can read the room but cannot ship code. Either way, the deployment stalls.

Automation trajectory. Resistant. The reading-the-room work, the exec-decision work, the shipping-under-political-pressure work is irreducibly human at this horizon. Agents augment FDEs (research, code generation, documentation) but cannot replace them.

Specialization. Vertical specialization is the dominant emerging pattern. OpenAI's vertical FDE org (Life Sciences, Semiconductor, Government) signals where customers are heading. Each vertical has its own scarcity profile and compensation band — Defense and Healthcare command the highest premiums.

Compensation reality (2026). Levels.fyi data: average FDE TC $238K, range $205-486K, Staff clearing $630K+. Palantir / OpenAI / Anthropic FDE: $350-550K mid-to-senior TC. UK FDE: £138K average, range £108-186K, top £253K+. Palantir's London FDE function sits in the £155-195K base band at senior levels with significant equity. New York has surpassed San Francisco as the FDE hub (35% vs 11% of postings).

Adjacent roles. A1 (junior version). A6 (Deployment Strategist — the FDE's structural counterpart in the paired model). D2 (Agent System Architect, designs what FDE then builds).

A3. AI Deployment Triage Specialist 🔴

What they do. Called in when an AI deployment has failed, partially failed, or is producing unintended consequences. Fixed window, usually 30-90 days, to either rescue or shut down responsibly. Post-failure work, structurally distinct from greenfield deployment. Diagnose what went wrong, stop the bleeding, identify whether the system is salvageable, restore to operation or document the lessons and decommission cleanly.

What good looks like. Starts with the operational impact, not the technical architecture. Asks "what's broken in your daily operations because of this system" before "what does the architecture look like." Has done this multiple times. Unflustered by political situations where someone is going to be embarrassed by the diagnosis. Writes clear post-mortems.

Demand signals. Companies that publicly announced an AI pilot 6-12 months ago and have gone quiet. Earnings calls where AI initiatives are being de-emphasized after being heavily promoted. Internal job postings for "AI Recovery Lead" or "Senior AI Engineer to assess existing systems." MIT's 95%-of-GenAI-pilots-fail finding and Pluralsight's 65%-of-organizations-have-abandoned-AI-projects finding make this a structural growth role. Pertama Partners 2026 data: 42% of companies abandoned AI initiatives in 2025; failed projects cost average $4.2M-$8.4M depending on failure mode.

Supply signals. Senior engineers and architects who have been through one or more AI deployment failures, learned from them, and can articulate the failure pattern. Often the people who built the failed system in the first place at a previous company. Public writing about "what went wrong with X" is a strong signal. Background in incident response, SRE, or post-mortem discipline transfers strongly.

Common failure when the wrong person is hired. A company brings in a generalist consultant who delivers a strategy deck. Or they bring in the original vendor of the failed system, who has motivated reasoning to keep it running. Or they bring in someone who tries to rebuild from scratch when triage was the right move.

Automation trajectory. Resistant. Crisis management, organizational diagnosis, "stopping the bleeding" require human judgment under uncertainty in politically loaded contexts.

Adjacent roles. B1 (AI Operations Engineer, ongoing reliability vs. acute crisis). C5 (AI Incident Responder, when failure produces regulatory or reputational consequences).

A4. Agent Workflow Implementer 🟡

What they do. Take a designed agent workflow and build it to production. Live in CrewAI, LangGraph, AutoGen, Mastra, the Vercel AI SDK, n8n with AI nodes. Handle the messy parts: state management, retry logic, error handling, tool authorization, observability hooks, cost monitoring. Make agentic systems run reliably under load.

What good looks like. Has shipped at least three production agent systems and can describe specifically how each one fails and how they instrumented for it. Paranoid about agent loops, infinite retries, cost spikes, and prompt drift. Writes defensive code. Builds in kill-switches by default.

Demand signals. Companies whose AI roadmap has moved from "single LLM call" to "multi-step agent." Postings asking for LangGraph, AutoGen, or CrewAI experience specifically. Vendor partnerships announcing "agentic" features. Any SaaS company under pressure to ship "agent mode." Gartner: 1,445% surge in enterprise inquiries about multi-agent orchestration in 2025.

Supply signals. GitHub contributions to agent frameworks. Public projects deploying agents in production. Show HN posts about specific agent architectures. Discord and forum activity in CrewAI, LangGraph, Mastra communities.

Common failure when the wrong person is hired. Backend engineer who has read the LangGraph docs and ships an agent that loops infinitely on edge cases or burns through API credits in a weekend. Or prompt engineer who can write good prompts but cannot reason about distributed system reliability.

Automation trajectory. Partial. A team of 3-4 implementers becomes 1-2 plus an agent fleet within 18 months. Routine implementation (boilerplate, tests, simple integrations) collapses to agents. Architecture decisions, edge-case debugging, and reliability instrumentation remain human.

Adjacent roles. B2 (Agent Reliability Engineer, maintains in production). D2 (Agent System Architect, designs the multi-agent structure).

A5. Agent System Integrator 🔴

What they do. Build, package, deploy, and successfully transfer an agent or agent system to a customer such that the customer can own and operate it after delivery. The deliverable is a working system, integrated into the customer's actual environment, with the customer's team trained, with documentation, evaluations, monitoring, and a clean handoff. The customer ends up with an asset they own, not a dependency on the integrator. This role exists specifically because every other delivery mode in the market — consultant, FDE, operator, vendor — leaves the customer somewhere short of full ownership of a working system.

What good looks like. Has shipped at least three working agent systems into customer environments where, six months later, the customer is still using the system without the integrator's involvement. This is the only signal that matters; most "AI deployments" cannot pass this test. Builds for transfer from day one. Documentation, evaluations, monitoring hooks, runbooks, training materials produced during the build, not as an afterthought. Has opinions about scope. Refuses to deliver a system the customer cannot operate, even if the customer is willing to pay. Right-sizes the deliverable. Trains the customer's team during the build, not at the end.

Demand signals. Companies burned by previous AI engagements where the consultant left and the system stopped working. SMBs and mid-market companies that explicitly want to own a working AI capability internally without permanently renting it. PE portfolio companies whose operating partner wants AI deployed at portfolio scale but needs each company to be operationally self-sufficient. The MSP market growing from $406B (2025) to $846B (2029) at 20.1% CAGR is the broader category; A5 specialists are the AI-specific entrants.

Supply signals. Practitioners with consulting backgrounds combined with engineering depth. Often founder-flavored — they have shipped products and know what "operable by someone who isn't me" actually requires. Public artifacts demonstrating clean handoffs: documentation, runbooks, training materials, evaluation frameworks built for previous deliveries.

Common failure when the wrong person is hired. A consultant who delivers strategy and a partial build, then leaves; the customer cannot operate what was delivered. Or a vendor whose generic product doesn't fit. Or an FDE who stays embedded and never transfers ownership.

Automation trajectory. Resistant. The transfer-of-ownership work — training the customer's team, building the operability layer, making judgment calls about what to include — is irreducibly human.

Adjacent roles. A1 (less ownership-focused). A2 (sustained embedded vs. transfer-and-leave). F-cluster operators (retain the fleet rather than transfer it).

A6. Deployment Strategist 🔴 (NEW in v0.3)

What they do. The non-engineering counterpart to A2 (FDE) and A5 (Agent System Integrator). Pioneered at Palantir, where the role is called "Echo" internally and where every customer engagement is paired: one Deployment Strategist + one or two FDEs. Salesforce uses the same model — explicit "pods" of one Deployment Strategist plus two FDEs locked onto a single customer for ~3 months. The Deployment Strategist scopes what gets built, wins stakeholder alignment, defines success criteria, and ensures the customer organization absorbs the work. Translates operational reality into technical scope, and translates technical output into operational decisions.

What good looks like. Generalist problem-solver who operates in environments where the problem is still forming and the client doesn't yet have language for what is wrong. Can impose structure without oversimplifying reality. Turns incomplete signals into a direction someone with real stakes is willing to follow. Sits with customer analysts to understand critical questions and locate biggest pain points. Identifies relevant datasets through deep engagement with workflows. Tailors workflows to the unique requirements of different user groups. Builds and delivers demos. Scopes potential engagements in new industries.

Demand signals. Direct: Palantir (the original DS role, called "Echo"), Salesforce (Deployment Strategist as part of their FDE pods), Anthropic (similar function in Applied AI Architect roles), OpenAI (similar function in Forward Deployed pre-sales; the Tomoro acquisition explicitly brought "Forward Deployed Engineers and Deployment Specialists" to the OpenAI Deployment Company). Indirect: any organization deploying AI at scale that has noticed FDEs alone don't close the customer-organizational-absorption gap. The Anthropic/Blackstone venture will need this role at high volume across PE portfolio deployments.

Supply signals. Ex-Palantir Deployment Strategist or "Echo." Strategy consultants from McKinsey/Bain/BCG who have moved into operating roles and learned data fluency. Former product managers from data-heavy environments. People with public writing that demonstrates ability to scope ambiguous problems. Track record of sitting in a room with both a CFO and an ML engineer in the same hour.

Common failure when the wrong person is hired. A traditional management consultant who can scope but cannot understand the technical constraints, leading to scope that the engineering team cannot deliver. Or a project manager who can manage scope but cannot win executive trust. Or an engineer promoted into the role who can talk technology but cannot scope the customer's actual operational problem.

Automation trajectory. Resistant. The role is fundamentally about reading rooms and scoping ambiguous problems. Both are at the far end of irreducibly-human work.

Compensation reality. Palantir: $110-170K base. Senior DS at scale-flavored ventures: $150-220K. Less than FDE because the technical depth is lower; the role is irreplaceable in the paired model.

Adjacent roles. A2 (FDE — the structural pair). A5 (Agent System Integrator — DS scope work feeds A5's transferable delivery). E1 (AI Implementation Lead — broader program-management flavor).

A7. Partner / Channel Solutions Architect 🔴 (NEW in v0.3)

What they do. Bridge between an AI lab/vendor and the partner ecosystem that ultimately delivers AI to end customers. Cultivate technical relationships with cloud partners (AWS, GCP, Azure) and non-cloud partners including Global System Integrators (GSIs) and Regional System Integrators (RSIs). Strengthen relationships with key partners to accelerate indirect revenue. Embed with partner technical teams to support troubleshooting and evangelize their company in the partner's developer community.

What good looks like. Strategic technical thought partner to the partnerships team. Deep understanding of partner landscape. Drives key strategic programs with partners. Validates and gathers feedback on products as they relate to use through partners, delivers feedback to relevant internal teams.

Demand signals. Direct: Anthropic (Partner Solutions Architect role), OpenAI (similar role in Applied AI; The OpenAI Deployment Company has Bain & Company, Capgemini, McKinsey & Company as explicit consulting and integration partners), every AI-native company at scale. Indirect: as the Anthropic/Blackstone venture and OpenAI/Bain venture scale, partner-channel architecture roles expand at every consultancy and integrator that joins the channel.

Supply signals. Ex-Solutions Architect at AWS, GCP, Azure. Ex-Partner Engineer at major SaaS companies. People with public presence in partner technical communities (re:Invent, Google Cloud Next, partner conferences).

Common failure when the wrong person is hired. Sales-heavy people who can manage relationships but cannot win technical credibility with partner engineers. Or technical people who cannot navigate the political dimensions of partnership.

Automation trajectory. Resistant. Partnership work is fundamentally human.

Adjacent roles. E1 (AI Implementation Lead — program-management flavor). A6 (Deployment Strategist — strategy flavor). E4 (Fractional Head of AI — when partnership work is fractional).

CLUSTER B — Reliability & Operations

B1. AI Operations Engineer 🟢

What they do. SRE, but for AI systems. Monitor model and agent behavior in production, detect drift, manage cost, track latency, instrument observability, set up alerting, handle incidents when AI systems misbehave. The on-call function for AI features. Increasingly distinct from traditional SRE because AI failure modes — hallucination, drift, cost explosions, tool-misuse, prompt-injection, jailbreaks — are unfamiliar to people coming from traditional infrastructure.

What good looks like. Has built dashboards for at least one production AI system. Has responded to a real production incident involving model behavior, not just infrastructure. Knows what "drift" actually looks like in their domain and how to detect it before users do. Has opinions about evals and uses them as production telemetry.

Demand signals. Any company running AI in production at non-trivial scale. Job posts for "AI Reliability Engineer," "AIOps Engineer," "ML Platform Engineer with LLM focus." Particularly hot at companies that have hit a public AI incident — once that happens, they realize they need this role permanently.

Supply signals. Background in SRE, DevOps, or infrastructure who have moved into AI. Contributors to evaluation tooling (Langfuse, PromptLayer, Helicone, Phoenix, Arize). People who have written publicly about AI incident response or production AI monitoring.

Common failure when the wrong person is hired. Traditional SREs who treat AI systems like deterministic infrastructure. Or ML engineers who can train models but have never operated one in production.

Automation trajectory. Collapsible. Within 12-18 months, one human supervises an agent fleet that does the routine monitoring, alert triage, dashboard creation, and incident-runbook execution.

Adjacent roles. A4 (Agent Workflow Implementer, builds; B1 operates). B2 (Agent Reliability Engineer, agent-specialist version). B4 (Inference Reliability, infrastructure-layer version).

B2. Agent Reliability Engineer 🟢

What they do. Specialist version of B1 focused on the unique reliability challenges of agentic systems specifically: multi-step failure cascades, tool-call errors, prompt injection in production, agent loops, cost-per-task variance, infinite retries, MCP server failures, OAuth drift, model fallback strategies.

What good looks like. Can describe the three most common ways an agent fails in production and how they would instrument for each. Has run a multi-agent system in production and can talk about specific failure incidents. Deeply skeptical of agent demos and asks specifically about edge cases.

Demand signals. Companies running multi-step agentic workflows in production, particularly customer-facing. Postings explicitly mentioning agent reliability or agent observability. Tooling vendors (Datadog, Helicone, Langfuse, Arize) hiring practitioners.

Supply signals. Public writing about agent failure modes. Contributors to evaluation frameworks. People who have given talks about "what we learned running X agent in production."

Common failure when the wrong person is hired. Traditional ML reliability engineers who don't understand the agent-specific failure surface. Or agent enthusiasts who can build but have not yet operated.

Automation trajectory. Collapsible. Self-monitoring agent fleets are already shipping. One human supervises a meta-monitoring agent layer within 12-18 months.

Adjacent roles. B1 (broader AI Ops). A4 (the implementation side). C5 (when reliability failures cross into regulatory or reputational territory).

B3. AI Cost & Capacity Operator 🟢

What they do. Specialist focused on the economics of running AI in production. Token cost optimization, model routing (small model first, large model fallback), batch vs. real-time decisions, caching strategies, capacity planning, vendor negotiation. Part FinOps, part platform engineering, part vendor management.

What good looks like. Has reduced an AI bill by 40-70% at a previous company without degrading user experience. Has opinions about which models to use for which tasks based on cost-quality tradeoffs. Builds automated routing layers. Negotiates with vendors and knows the actual unit economics.

Demand signals. Any company whose AI bill has crossed $50K/month. Job posts for "AI FinOps," "AI Cost Engineer," "ML Infrastructure Optimization." Forrester: 25% of 2026 enterprise AI spend being deferred to 2027 because of ROI pressure.

Supply signals. FinOps practitioners who have moved into AI. Engineers who have written public posts about cost optimization at specific companies.

Automation trajectory. Collapsible. Cost optimization is fundamentally pattern-matching across usage data; agent fleets do this well. One human + cost-optimization fleet replaces a 3-4 person team within 18 months.

B4. AI Inference & Model Serving Reliability Engineer 🟡 (NEW in v0.3)

What they do. Distinct from B1, B2, B3. Works at the layer between the model and the application — inference latency, throughput, multi-region deployment, model serving infrastructure, GPU fleet utilization, multi-cloud cost normalization. The role exists at every hyperscaler (AWS, Azure, GCP), every AI infra startup (Modal, Replicate, Together AI, Anyscale, Crusoe), every frontier lab, and increasingly at enterprises running their own GPU fleets.

What good looks like. Has built and operated production inference infrastructure at meaningful scale. Has opinions about model serving frameworks (vLLM, TGI, TensorRT, etc.). Builds observability for accelerator fleets. Optimizes compute efficiency at the hardware level.

Demand signals. Direct: Anthropic ("Staff + Sr. Software Engineer, AI Reliability"; "Software Engineer, Cloud Inference"; "Software Engineer, Cloud Inference Launch Engineering"; "Software Engineer, Inference Deployment"; "build and operate production data pipelines and observability systems for Anthropic's accelerator fleet"). OpenAI, Google, Meta all hire for this. Hyperscalers, AI infra startups (Modal, Together AI, Anyscale), Crusoe.

Supply signals. ML infrastructure engineers from frontier labs or AI-native companies. Backgrounds in distributed systems with ML specialization. Public contributions to inference frameworks.

Common failure when the wrong person is hired. Generic ML engineers without infrastructure depth. Or distributed systems engineers without ML model knowledge.

Automation trajectory. Partial. Some automation of routine ops, but the architectural and capacity-planning judgment work remains human at this horizon.

Compensation reality. $200-400K+ TC at frontier labs and hyperscalers, with significant equity at AI-native infrastructure startups.

Adjacent roles. B1 (application-layer reliability). B3 (cost-focused). D2 (architectural overlap).

CLUSTER C — Governance, Risk & Compliance (Summary)

Cluster C overlaps significantly with Part III (the Compliance Layer) below. The roles described here are the ones that overlap most with general AI workforce concerns. Deeper compliance-specific roles are described in Part III.

C1. AI Audit & Conformity Lead 🟡

What they do. Make AI systems defensible — to auditors, regulators, boards, and customers. Build the documentation, audit trails, evaluation records, model cards, data lineage, and human-in-the-loop guarantees required for compliance with EU AI Act, NYC Local Law 144, Colorado AI Act, emerging US state laws, SOC 2 controls for AI, and customer-driven AI risk assessments.

What good looks like. Has shipped at least one conformity assessment under a real regulatory framework. Can read EU AI Act Annex III and tell you in 60 seconds whether a given system is high-risk. Builds governance into the engineering process from day one rather than retrofitting it.

Automation trajectory. Partial. Documentation, audit-trail generation, and compliance reporting collapse to agents. Signed-name attestation, regulatory-judgment calls, and cross-functional negotiation work remain human.

C2. AI Risk & Policy Analyst 🔴

What they do. Define what AI can and cannot be used for inside an organization. Write the AI policies, define data use boundaries, set approval workflows. WRITER survey: 67% of executives report data breaches from unapproved AI tools.

Automation trajectory. Resistant. Policy is fundamentally a human-decision artifact under ambiguity.

C3. Model & Vendor Governance Manager 🟢

What they do. Manage the ongoing relationship between an organization and its AI vendors. Track which models are in use, monitor vendor terms changes, negotiate enterprise agreements, manage data processing agreements.

Automation trajectory. Collapsible. Vendor monitoring, contract change tracking, comparison benchmarking are agent-suited tasks.

C4. AI Agent Steward 🔴

What they do. Monitor deployed agents for ethical, behavioral, and operational compliance. Distinct from B1/B2 because the focus is on behavior, not reliability. Reviews agent logs, investigates flagged behaviors, manages incident response when agents do something problematic.

Automation trajectory. Resistant. The human who oversees agent behavior cannot themselves be an agent (trust regression).

(For C5-C9, see Part III.)

CLUSTER D — Design & Architecture

D1. AI Workflow Designer 🟡

What they do. Redesign business processes so AI can reliably assist, automate, or augment them. Not an agent designer — they don't build the agent. They design the work the agent does: intake forms, ticketing templates, handoff protocols, approval steps, exception flows. Reduce ambiguity in workflows so AI outputs are easier to validate and use.

What good looks like. Has shipped at least three workflow redesigns where AI was integrated successfully. Starts by mapping the existing workflow in detail, including the informal parts. Understands that 60-80% of AI failure is workflow ambiguity, not model capability.

Demand signals. Operations-heavy companies deploying AI. Postings for "AI Workflow Lead," "Process Designer with AI focus." Often hidden inside broader operations or transformation roles. Particularly hot at insurance, healthcare, financial services, professional services.

Supply signals. Operations professionals who have learned AI. Six Sigma / Lean practitioners moving into AI integration. Process designers from BPO who picked up AI fluency.

Automation trajectory. Partial. Workflow mapping and documentation collapse to agents. Redesign judgment, especially for politically loaded cross-functional processes, remains human.

D2. Agent System Architect 🔴

What they do. Design multi-agent systems at the architectural level. Decide which tasks should be agents vs. functions, design tool-call boundaries, architect handoff protocols between agents, design memory and state management at the system level.

What good looks like. Has architected at least one multi-agent system that ran in production. Can articulate trade-offs in agent system design (single big agent vs. multiple specialized agents, dynamic vs. static tool sets, memory vs. retrieval). Writes architectural decision records.

Demand signals. Companies moving from single-agent to multi-agent systems. Postings for "Agent Architect," "Multi-Agent System Designer," "Senior AI Engineer (Architecture)."

Supply signals. Software architects who have moved into AI. Researchers who have shipped production. Conference speakers on multi-agent systems. GitHub maintainers of agent frameworks.

Automation trajectory. Resistant. Architectural judgment under uncertainty across multiple stakeholders is the kind of work that gets harder, not easier, with more agents in the loop.

D3. Prompt and Context Engineer 🟢 (renamed in v0.3)

What they do. Design and maintain the context layer for AI systems — what information the model sees, in what order, structured how. Includes RAG architecture, prompt engineering at the systems level, document chunking strategies, retrieval ranking, context window management, agent skills systems (SKILL.md, sub-agent context patterns), and the bridging-layer between model capability and product experience. Anthropic explicitly hires for this as "Prompt Engineer, Agent Prompts & Evals" and "Prompt and Context Engineers."

What good looks like. Has measurably improved an AI system's quality by improving its context layer rather than its model or prompt. Runs experiments. Builds evaluation harnesses that detect context-quality regressions. Has a structured methodology for context design. From Anthropic's own JD: bridges "the gap between model capabilities and real product experience, working with product teams to build consistent, safe, and beneficial user experiences across all product surfaces."

Demand signals. Direct: Anthropic, OpenAI, every frontier lab and AI-native scaleup. Indirect: any company whose AI quality has plateaued and they've already tried bigger models. Companies adopting RAG at production scale. Postings for "Context Engineer," "Senior Prompt Engineer," "RAG Engineer." Demand grew 135.8% in recent quarters per Coursera salary data; projected CAGR 32.8% through 2030.

Supply signals. Public writing about specific RAG implementations and what they learned. Contributors to retrieval libraries. People who have shipped multiple measurable context improvements at named companies.

Compensation reality. $100-250K typical, top performers at OpenAI and Anthropic exceed $300K.

Automation trajectory. Collapsible. Context experimentation, chunking optimization, retrieval ranking, eval harness running are largely agent-runnable. Human designs methodology; agents execute experiments.

D4. Human-AI Handoff Designer 🔴

What they do. Design the moments where AI hands off to humans and vice versa. Where does the agent escalate? What does the human see? What information needs to be preserved? How does the human's decision feed back into the agent? Most production AI systems are not fully autonomous — they are human-AI hybrid systems, and handoff design determines whether the hybrid works.

What good looks like. Has shipped at least two production human-AI workflows where the handoff design measurably improved outcomes. Understands both UX and AI behavior. Designs for the human's cognitive load, not just the agent's capability.

Demand signals. Customer support orgs deploying AI. Healthcare and legal contexts where humans must remain in the loop. Postings for "AI UX Designer," "Human-AI Interaction Designer."

Supply signals. UX designers who have moved into AI. Conversation designers from chatbot era who have leveled up.

Automation trajectory. Resistant. Designing for human cognitive load requires humans who understand humans.

D5. AI Evaluations Engineer 🟡 (NEW in v0.3)

What they do. Design eval suites that measure model and agent quality, safety, and reliability before and after deployment. Build evaluation harnesses, run statistical analyses on eval results, identify gaps in eval coverage, productionize evals into model-training and deployment pipelines. The role bridges research methodology and production engineering. From Anthropic's own description of the equivalent role: "How do we measure whether a model is safe? How do we create evaluations that reflect real-world usage rather than synthetic benchmarks? How do we know our graders are accurate?"

What good looks like. Has built at least one eval suite that influenced a real shipping decision. Understands statistical methodology (Anthropic's published "A statistical approach to model evaluations" paper is foundational). Can decompose an eval failure into model issue vs. eval design issue. Productionizes evals as monitoring telemetry, not one-off measurements.

Demand signals. Direct: Anthropic ("Research Engineer, Model Evaluations"; "Applied Safety Research Engineer, Safeguards"), OpenAI, Google DeepMind, every frontier lab. The UK AI Security Institute (AISI) uses Anthropic's open-source Petri tool to evaluate models for sabotage propensity. Indirect: any company shipping AI to production at scale; increasingly any company maturing its AI governance.

Supply signals. Public contributions to evaluation tooling (Inspect, Petri, Scout). Authored papers on eval methodology. Contributors to eval frameworks (LangSmith, Braintrust, Phoenix, Promptfoo). Statistics or experimental-design background plus engineering.

Common failure when the wrong person is hired. Generic ML engineers without eval-design depth, who produce "vibe-based evals" (OpenAI's term for the anti-pattern). Or statisticians without engineering depth, who can design but cannot productionize.

Automation trajectory. Partial. Eval execution and routine grading collapse to agents (Anthropic is already running Automated Alignment Researchers as a research project). Eval methodology design remains human at this horizon.

Adjacent roles. D3 (Prompt and Context Engineer — overlap on production iteration). C-Research roles in Part III (overlap on safety evals). Alignment Researcher in Part IV (overlap on research-flavored eval work).

CLUSTER E — Translation & Enablement

E1. AI Implementation Lead 🔴

What they do. Run AI rollouts inside an organization the way someone used to run cloud migrations. Coordinate stakeholders, define success metrics, manage vendor selection, sequence the rollout, ensure AI integrates cleanly with existing tech and process. Distinct from A1/A2 because they don't build — they orchestrate.

What good looks like. Has run a major technology rollout to completion (not necessarily AI). Bilingual between business and technical. Makes hard prioritization calls. Understands AI adoption is 80% organizational change and 20% technology.

Demand signals. Mid-market and enterprise companies whose AI strategy has moved beyond pilots. Postings for "AI Program Manager," "AI Implementation Lead," "Director of AI Adoption."

Supply signals. Program managers who have run major technology initiatives. Often coming from cloud migration, ERP rollout, or digital transformation work.

Automation trajectory. Resistant. Orchestrating organizational change is human work.

E2. AI Enablement Trainer 🟢

What they do. Build the internal training, playbooks, prompt libraries, usage standards, and "how to work with AI" capability inside an organization. The bottleneck for most enterprises adopting AI is not access — it is effective use.

What good looks like. Has trained at least 100 people on AI tools and can describe what worked and what didn't. Builds artifacts that scale rather than relying on 1:1 enablement. Has a methodology for measuring AI fluency, not just AI awareness.

Demand signals. Companies that have rolled out AI tools (Copilot, ChatGPT Enterprise, Claude for Enterprise, Glean, Notion AI) and discovered usage is low or inconsistent. Postings for "AI Enablement Lead," "AI Adoption Manager," "Director of AI Training." Particularly hot at companies with $1M+ AI tooling budget but no productivity gain to show. Cognizant Skillspring (April 2026 launch) directly addresses this market.

Supply signals. L&D professionals who have moved into AI. Internal evangelists who have led AI rollouts at named companies.

Automation trajectory. Collapsible. Training material generation, playbook drafting, Q&A delivery can be largely agent-driven.

E3. AI Translator 🔴

What they do. Sit between business stakeholders and technical AI teams, translating in both directions. Distinct from product management because the translation problem is denser and more frequent.

What good looks like. Has shipped projects where their translation was the reason the project succeeded. Understands both technical and business sides at depth. Adjustable in altitude — can talk to a CFO and an ML engineer in the same hour.

Demand signals. Mid-market and enterprise companies running multiple parallel AI initiatives. Postings for "AI Product Manager (the new flavor)," "AI Solutions Lead."

Supply signals. Product managers who have shipped AI features at scale. Strategy consultants who have moved into operating roles.

Automation trajectory. Resistant. The translation problem is the human's reason for being.

E4. Fractional Head of AI 🔴

What they do. Senior AI leadership, on a fractional basis (typically 8-32 hours per week, $5-20K+/month). Sets AI strategy, sequences initiatives, hires and manages AI teams, reports to board on AI progress. Increasingly common at companies that need senior AI leadership but cannot justify a full-time hire.

What good looks like. Has led an AI function at a real company before. Senior enough to be in board conversations and trusted enough to make hiring decisions. Honest about what AI can and cannot do at the company's stage. KORE1 reports the typical engagement runs 3-4 weeks scoping-to-signed, with the role often converting to interim Chief AI Officer if AI budget exceeds $1M annually.

Demand signals. Companies that have outgrown "the CTO does AI on the side." Companies whose AI budget is starting to require senior oversight. PE portfolio companies whose operating partner is pushing AI adoption.

Supply signals. Former heads of AI at named companies who are now consulting. Senior practitioners with public writing and speaking presence. Alumni of frontier labs or AI-native startups who have moved to fractional work.

Automation trajectory. Resistant. Board-room presence cannot be agentized.

A note before this map

What follows is not a list of jobs. It is a description of a new economic unit that emerged in the last 18 months and does not fit cleanly into any prior category of work. The operator is not an employee. Not a freelancer. Not a consultant. Not an agency. Not a SaaS founder. Not a fractional executive. They share traits with each but are none of them.

Three things became simultaneously possible in 2025-2026 that had never been possible together before: a single human can curate, evaluate, and supervise a portfolio of agents reliable enough to produce real customer-facing output; the cost of running that portfolio is structurally trivial ($300-500/month for a serious stack vs. $80-120K/month for an equivalent team); the market for agent-produced output is real and paying.

Real examples already operating at scale: Pieter Levels at $3M+ ARR solo. Danny Postma at $300K/month with HeadshotPro. Sarah Chen at $420K in 8 months as solo AI design agency. Maor Shlomo's Base44 sold to Wix for $80M in 6 months. Anthropic's Amodei: 70-80% probability of first one-person billion-dollar company in 2026. NVIDIA internally: 100 AI agents per human (7.5M agents serving 75K humans). Jensen Huang at GTC 2026: "In the future, the IT department of every company is going to be the HR department of AI agents."

But — and this is critical — the operator pattern is also visible at scaleup level. Cursor (Anysphere) at $2B ARR with 60 employees is the operator pattern at scaleup scale. Midjourney at $500M revenue with ~107 employees is the operator pattern at scaleup scale. The pattern isn't just solo. The pattern is small humans + many agents, at every size of company.

What follows is the first attempt to name the five operator types I see operating in the field today.

The five operator types

F1. The Solo Agent Operator

What they are. One human, one curated agent fleet, several customers. The dominant emerging pattern.

Real examples. Pieter Levels (Nomad List, RemoteOK, PhotoAI, $3M+ ARR). Danny Postma (HeadshotPro, $300K/month). Sarah Chen (AI design agency, $420K in 8 months). Hundreds of others operating quietly under traditional framing.

What good looks like. A small portfolio of customers (3-15 typical) each paying recurring fees ($1-15K/month). Fleet is tuned — evals running, failure modes known, human-in-the-loop checkpoints where it matters. Obsessive about output quality because reputation is the product. Publishes (writing, build-in-public, demo videos) because that's how they're discovered.

Demand signals. SMBs and mid-market companies who want specific output but cannot afford an in-house team and don't trust generic SaaS to deliver it. Founders who need leverage but cannot or will not raise capital to hire. Marketing-heavy businesses where output volume matters more than deep relationship.

Operator signals (recognizing a real one). Charges recurring fees, not hourly. Revenue not bounded by hours. Portfolio of customers, not a single client. Publishes artifacts demonstrating they understand the unit. Has stories of failures and recoveries.

Pricing. $1-3K/month narrow output; $3-8K/month substantial operational coverage; $8-25K/month embedded operators.

Adjacent units. F2 next size up. F3-F5 specialized variants.

F2. The Boutique Agent Operator

What they are. 2-5 humans running multiple specialized agent fleets for multiple customers. Small-agency pattern adapted for the agent era.

Real examples emerging in 2026. Bug0 (described publicly: "Outcome-as-a-Service" with two tiers — Studio for self-serve AI test generation, Managed for full FDE pod that owns customer QA). Multiple boutique operators in legal-tech, healthcare-ops, financial research.

What good looks like. Systematized the parts of fleet operation that can be (evals, monitoring, customer onboarding, billing) and kept human-decision in the parts that can't (fleet design, customer relationship, judgment on quality). Scales by adding fleets, not headcount. Specialized — generalist boutiques struggle.

Pricing. $5-25K/month typical, premium engagements $50K+/month for substantial multi-fleet coverage.

F3. The Vertical Agent Operator

What they are. Operator (solo or boutique) specialized in one industry, running fleets tuned to that industry's specific data, workflows, regulatory environment.

Real examples emerging. Legal (document review, contract drafting), real estate (BatchData and Cotality MCP integrators), healthcare (intake fleets, scheduling fleets), accounting (Mercury MCP integrators adapted by solo CPAs).

What good looks like. Speaks the industry's language fluently. Has shipped fleets that handle the industry's specific edge cases. Customer references within the industry. Understands the regulatory environment for that industry specifically.

Pricing. Often higher than F1/F2; $10-50K/month typical for substantive vertical engagements.

F4. The Function Agent Operator

What they are. Operator specialized in one cross-industry function (sales operations, customer support, research, recruiting, financial analysis, content production).

Real examples. Sales-ops fleet operators serving B2B SaaS. Support fleet operators serving e-commerce. Research fleet operators serving PE firms.

What good looks like. Has shipped fleets for the same function across multiple industries. Has clear methodology refined across customers. Has position on canonical metrics for the function.

Pricing. Similar to F3. Outcome-based pricing emerges in F4 because metrics are clear (per-deal, per-ticket, per-research-output).

F5. The Integration Agent Operator

What they are. Operator running agents that integrate other people's tools and data on behalf of customers. The plumber pattern. Maintains the connective tissue between customer's existing systems and the AI capabilities they want.

What good looks like. Has shipped reliable integration layers for at least three customers. Opinions about MCP, OAuth, observability, vendor reliability. Has evals and monitoring for integration quality, not just output quality.

Demand signals. Companies whose AI projects are stalling because the integration layer is fragile. Companies with multiple AI vendors who need someone to manage the combined surface.

Pricing. Recurring monthly fees ($3-15K typical), often per-integrated-system or per-active-fleet pricing.

The Founding Engineer at AI-native company — a polymath archetype

A note worth flagging: at AI-native companies under 50 people (Cursor, Midjourney, Anysphere, early-stage everywhere), the Atlas roles aren't separate hires. They're aspects of one polymath role. The founding engineer ships features, embeds with strategic customers (FDE work), tunes prompts (D3), runs evals (D5), manages production reliability (B-cluster). Cursor's culture description: "ship-fast, eng-driven, product-impact, flat hierarchy, many-hats." The Atlas describes the specializations but recognizes that early-stage AI-native companies don't separate them.

This affects shipstacked's matching engine: the unit-of-supply varies dramatically by company stage. SMB / mid-market customers want specialists. AI-native scaleups under 50 people want polymaths.

How operators relate to companies — the engagement model

Customers do not hire operators. The verb is wrong. They engage them.

• Operator retains the fleet. Customer rents the output.
• Pricing is recurring, not project-based.
• Scope is defined by output, not hours.
• Trust is at the founder/operator level.
• Termination is graceful, not catastrophic.

This engagement model does not fit cleanly into existing procurement, legal, or HR frameworks at most companies. That mismatch is one of the friction points slowing operator adoption at large enterprises and one of the reasons SMBs and mid-market companies are adopting operators faster.

Why compliance gets its own part

The buyer is different (Chief Compliance Officers, General Counsels, Chief Risk Officers, Heads of Trust & Safety). The calendar is concrete (EU AI Act August 2 2026, NYC LL144, Colorado AI Act, sector-specific frameworks). The unit economics make this the highest-margin part of the agentic-economy labor surface — fines run to 7% of global revenue under EU AI Act.

Three sub-clusters

The Atlas v0.3 restructures the compliance layer into three sub-clusters reflecting how frontier labs actually organize this work. Anthropic alone has at least nine distinct safety/policy specialisms across these three areas.

Sub-cluster C-Research (Frontier Red Team flavor)

Research-flavored, future-risks-focused. Lives inside Policy organizations at frontier labs. Anticipates 6-18 month risks. Publishes. Shapes regulation.

C2. AI Risk & Policy Analyst — (also in Cluster C summary). Define AI policy at the org level.

C6. AI Red Team Lead 🔴. Adversarial testing of AI systems. Splits at frontier labs into multiple sub-specialties:

• Frontier Red Team (Autonomy) — Anthropic specifically: build and eval model organisms of autonomous systems and develop defensive agents.
• Frontier Red Team (Cyber) — anticipating "expert-level, even superhuman" AI capabilities in cybersecurity domains.
• Frontier Red Team (Emerging Risks) — societal-scale risks, novel risks emerging from agents interfacing with the external world.

Demand signals: Anthropic Frontier Red Team across multiple sub-orgs (Autonomy, Cyber, Emerging Risks). OpenAI, Google, Meta equivalents. UK AI Security Institute (AISI). US AI Safety Institute. Increasingly at non-AI-lab companies that deploy AI at scale.

Supply signals: Public CTF participation in AI-specific challenges. Contributions to adversarial-AI research. Practitioners with backgrounds at frontier labs or security-conscious AI-native companies. Track record in offensive security research, vulnerability research, exploit development. Research or professional experience applying LLMs to security problems.

Sub-cluster C-Operations (Safeguards flavor)

Operations-flavored, current-product-protection. Handles deployed-system risks in real time. Lives inside product or trust-and-safety organizations.

C5. AI Incident Responder 🔴. Specialist who handles AI incidents after they happen. Prompt injection in production. Data exfiltration via agent. Hallucination causing real-world harm. Agent making unauthorized commitments. Closer to security incident response than reliability engineering.

Demand signals: Direct: Anthropic (Incident Manager, Detection & Response team, "leading incident response efforts and driving systemic improvements post-incident"). Companies running customer-facing AI at scale. Companies in regulated industries where AI incidents trigger regulatory disclosure.

Supply signals: Security incident response practitioners who have moved into AI. People who have publicly written about specific AI incidents.

C8. AI Procurement & Vendor Risk Assessor 🔴. Pre-purchase risk assessment of AI vendors and AI-enabled SaaS products. Distinct from C3 (ongoing relationship) — C8 is the intake role.

What good looks like: AI-specific vendor assessment framework that goes beyond generic SaaS procurement. Can read a model card and identify gaps. Negotiates AI-specific terms (training-data use, output ownership, model-update notification, incident disclosure obligations).

Anthropic Safeguards Red Team (sub-specialty). Distinct from Frontier Red Team — focuses on current product abuse: coordinated account manipulation, payment fraud, novel exploitation of product features. Anthropic posts: "Red Team Engineer, Safeguards" and "Staff Red Team Specialist, Safeguards."

Sub-cluster C-External (Trust & Safety flavor)

Partnership-driven, expert-coordination-flavored. Works with external subject-matter experts on policy-specific risks.

C7. Data Provenance & Training-Data Compliance Officer 🟡. Track what data was used to train, fine-tune, or RAG-feed a model. Manage copyright and licensing exposure. Handle data subject access requests for AI systems. Manage right-to-be-forgotten in models.

What good looks like: Can produce a complete data lineage trace for any production AI system on demand. Understands the IP exposure of training-data choices. Has opinions about which licensing frameworks (CC, OpenRail, custom commercial) work for which use cases.

C9. Vulnerable User Protection Lead 🔴. Specialist focused on AI systems' interactions with minors, healthcare patients, vulnerable populations. The legal exposure here is enormous — children's safety frameworks, patient-protection regulation, accessibility law, all apply to AI systems and most companies have not adapted their AI deployments.

Demand signals: Any company whose AI-powered products touch minors, healthcare patients, or vulnerable populations. EdTech, healthtech, fintech consumer products, AI-powered customer service touching vulnerable callers.

Policy Vulnerability Testing (PVT) Coordinator. Anthropic's own framing for the role that coordinates external subject-matter experts (Thorn for child safety, Institute for Strategic Dialogue on election integrity, Global Project Against Hate and Extremism on radicalization). This is a distinct role: policy expert who orchestrates external testing partnerships.

Why this gets its own part

Alignment and interpretability research is a distinct labor market. Different career path than compliance. Different supply pool (academic ML researchers, philosophy/cognitive-science backgrounds, safety engineering). Different demand pool (frontier labs primarily, AI safety institutes, governments, well-funded safety nonprofits). Different funding mechanisms ($200M+ in safety grants flowing through dedicated programs in 2026 per Granted AI's data).

Anthropic alone has dedicated teams for: Alignment, Interpretability, Societal Impacts, Frontier Red Team. Their published research output is substantial: Petri 3.0 (open-source alignment auditing tool, used by UK AISI), Automated Alignment Researchers, alignment audits, alignment faking research, scalable oversight, mechanistic interpretability.

The roles

Alignment Researcher

What they do: Research the question of how to ensure AI systems pursue goals compatible with human values. Reward specification, scalable oversight, RLHF, AI control, model organisms of misalignment. Anthropic's Alignment team works "to understand the risks of AI models and develop ways to ensure that future ones remain helpful, honest, and harmless."

Demand signals: Anthropic, OpenAI, Google DeepMind, Redwood Research, ARC, MIRI. Anthropic Fellows Program (May and July 2026 cohorts) is the leading pipeline — over 80% of fellows produce publishable research, over 40% subsequently join Anthropic full-time.

Supply signals: Academic ML researchers with safety focus. PhDs in AI safety, alignment, or related fields. Alumni of MATS (ML Alignment Theory Scholars), Redwood Research Residency, ARC research positions, Anthropic Fellows. Public output (papers, alignment forum posts, GitHub contributions to alignment tools).

Interpretability Researcher

What they do: Discover and understand how large language models work internally. Mechanistic interpretability — circuits, superposition, feature visualization. Build attribution graphs that partially reveal the steps a model took internally to decide on a particular output. Anthropic: "AI models like Claude talk in words but think in numbers."

Demand signals: Anthropic Interpretability team. Google DeepMind Interpretability. Redwood Research. EleutherAI. AI Safety Institutes (UK AISI, US AISI, others emerging). Academic labs at top CS programs.

Supply signals: Public research output (papers, interpretability tools on GitHub). Backgrounds in neuroscience, math, theoretical CS, physics. Often academic before industry.

Model Behavior Researcher

What they do: Study how models behave in the wild. Why do they hallucinate? Why do they fake alignment? Why do they engage in reward hacking? Anthropic's Societal Impacts team explores how AI is used in the real world.

Demand signals: Frontier labs. AI safety nonprofits. Government safety institutes.

Safety Evaluation Researcher

What they do: Designs evaluations specifically for safety properties (vs. capabilities). Anthropic's "Applied Safety Research Engineer, Safeguards" role: "How do we measure whether a model is safe? How do we create evaluations that reflect real-world usage rather than synthetic benchmarks?"

Adjacent to D5 (AI Evaluations Engineer) but research-flavored rather than production-flavored.

Talent Development Models

The supply pool for Part IV is largely created through structured programs:

• Anthropic Fellows Program: 4-month structured program, $15K/month stipend, mentor-led, project-based, public-output-focused. May and July 2026 cohorts open. Research areas: scalable oversight, adversarial robustness and AI control, model organisms of misalignment, mechanistic interpretability, AI security, model welfare. Operates in US, UK, Canada.
• MATS (ML Alignment Theory Scholars): Selective program pairing scholars with established alignment researchers for multi-month mentored projects. Alumni at Anthropic, DeepMind, Redwood Research.
• Redwood Research Residency: Hands-on residency focused on empirical alignment research.
• ARC research positions: Theoretical alignment research focus.

The Atlas notes these because the discovery layer for Part IV roles is fundamentally different from the rest of the Atlas — it runs through structured fellowship programs more than through traditional hiring.

Why this gets its own part

The model training and RLHF labor market is a distinct population that the rest of the Atlas does not capture cleanly. It's not employed work (most is contract). It's not operator work (the worker doesn't run an agent fleet — they train one). It's not freelance work in the traditional sense (the structure is more durable).

Already a $1B+ market dominated by Mercor ($1B annualized revenue in February 2026, paying $1.5M+ daily to a 300K+ contractor network). Scale AI built a $14B business on this before Meta acquired them. The market is structurally distinct from the customer-facing AI deployment work that is shipstacked's primary territory.

The Atlas describes Part V for completeness. shipstacked does not currently compete in this market (Mercor is the dominant player) but the domain-expert tier overlaps with shipstacked's vertical-specialist supply pool.

The three tiers

Mass-Market RLHF Contractors

Rate-and-rank work. Follow guidelines. Score AI responses on quality. ~$25/hr typical. Scale AI / Surge AI / Labelbox / Appen / Mercor all source from this pool. Global, ~300K+ practitioners.

Domain-Expert RLHF Specialists

Domain expertise applied to AI training. Lawyers training legal AI. Doctors training medical AI. Coders training code AI. Bankers training finance AI. ~$85+/hr at Mercor. The core of Mercor's $1B revenue. Top tier of the RLHF labor market.

This tier overlaps with shipstacked's vertical-specialist supply pool. A legal AI practitioner currently doing RLHF work for OpenAI could plausibly also do A1 (AI Integration Operator) work at a law firm. The same human, two different markets.

AI Quality Auditors / Red Team Contractors

Review other trainers' work. Establish guidelines. Test AI systems for harmful outputs. Often $120K+ FTE roles or premium contract rates ($150K+ equivalent). Live inside or adjacent to Cluster C compliance work in Part III.

Strategic note for shipstacked

Mercor's vulnerability moment (LiteLLM supply-chain breach March 2026, 4TB data exposure, class-action lawsuits, Meta paused work) creates a window. But entering this market requires building data-labeling infrastructure that competes with Mercor's $10B-valuation operational scale. shipstacked's better play is to focus on the adjacent opportunity: where domain-expert RLHF specialists are simultaneously customer-facing AI integration practitioners. That overlap is invisible to Mercor (focused on labs) and invisible to LinkedIn (no taxonomy for it).

Why this gets its own part

Domain practitioners who learned AI, rather than AI practitioners who learned a domain. Distinct supply pool. Distinct career path. Increasingly the largest segment of AI hiring globally per multiple staffing-market data sources:

• Healthcare AI: 640,000 positions in 2026, fastest CAGR (36.8%), reaching $110B+ by 2030.
• Manufacturing AI: 620,000 positions in 2026.
• Financial Services AI: 470,000 positions, top compensation in finance ($300K+ specialist roles).
• Defense / Government AI: Cloud & Defense AI boom in DC and Seattle, security clearance premiums.

Specialist domain experts command 30-50% higher salaries than generalists. Over 75% of AI job listings specifically seek domain experts.

The vertical-specialist roles

Healthcare AI Engineer / Specialist

Medical imaging, genomics, clinical deployment, regulatory writing, submissions, scientific operations. OpenAI's Life Sciences FDE JD: "workflows across discovery, clinical development, regulatory writing, submissions, or scientific operations where validation strategy, auditability, compliance constraints, and reviewer expectations shaped system design and rollout."

Companies: Navina, Tempus, Butterfly Network, Philips, Siemens Healthineers, Epic, every major hospital system.

Demand signals: HIPAA-compliant AI deployments. AI integration with EHR systems. Clinical decision support. Medical imaging analysis. Surgical robotics. Pathology automation.

Supply signals: Practitioners with medical background plus AI fluency. MD/PhD plus engineering. Health-tech engineering experience.

Legal AI Engineer / Legal Technologist

AI-native law firms (Harvey, Cuckoo, Spellbook, others). Document automation. Contract intelligence. Legal research automation. Regulatory compliance automation.

Demand signals: AI-native law firms hiring directly. Big Law AI initiatives. Legal tech startups.

Supply signals: JD plus engineering. Practicing lawyers who have learned AI. Legal-tech engineers with domain depth.

Financial Services AI Engineer

Fraud detection, algorithmic trading, risk assessment, regulatory compliance, KYC/AML automation.

Demand signals: JPMorgan, Goldman Sachs, every major bank. Fintech AI startups. Hedge funds (top compensation for trading-model engineers — $400K+ TC at hedge funds).

Supply signals: Quant background plus AI. Banking technology background plus AI fluency.

Defense / Government AI Engineer

OpenAI's Government FDE explicitly requires security clearance. Anthropic's Detection & Response. Palantir's government work. Anduril, Shield AI, Helsing.

Demand signals: Security clearance gating. Classified-environment deployment. Government acquisition cycles.

Supply signals: Security clearance plus AI fluency. Veterans transitioning into AI. Ex-Palantir government FDEs.

Compensation reality: Defense and Healthcare command the highest premiums in vertical FDE work per KORE1 data — $155-230K offers for forward-deployed engineers, with vertical mattering significantly.

Manufacturing AI Engineer

Quality control automation, predictive maintenance, robotics integration, supply chain AI, factory floor deployment.

Demand signals: Boeing, Tesla, every major manufacturer. Industry-specific AI startups.

Supply signals: Manufacturing engineering background plus AI. Industrial automation experience.

Strategic implications

Vertical specialization is more concentrated than generic AI hiring. The supply pool for each vertical is structurally smaller. Compensation premiums (30-50% above generalists) reflect this scarcity. shipstacked's matching engine should treat vertical as a first-class dimension across multiple Atlas roles, not as a tag added at the end.

If you are a company hiring, start with the symptom, not the role. Describe what is broken in your operations or what you are trying to ship. The right role, operator type, vertical specialist, or compliance hire will reveal itself from the symptom. If you start with the title, you will hire the wrong category 50% of the time.

If you are a practitioner, look at the work, not the title. Read each role's "what good looks like" sections honestly. Some of you are doing two of these simultaneously and your title doesn't reflect either.

If you are an operator, you may not have recognized yourself in any of the Cluster A-E roles, and that is the point. Part II is for you. The unit you are running has not had a name. Use the operator type that fits and begin describing yourself accurately to customers and to peers.

If you are an alignment or interpretability researcher, Part IV is your map. The career path runs through Anthropic Fellows, MATS, Redwood Residency, ARC, and direct hires at frontier labs. The discovery layer for your work is fundamentally different from the rest of the Atlas.

If you are a domain practitioner who has learned AI, Part VI is for you. You command 30-50% higher compensation than generalists with equivalent AI fluency. Your scarcity is structural; price accordingly.

If you are an investor, operator, or analyst, this Atlas is a map of the labor surface where the next generation of AI-implementation work is happening. The roles are real. The shortages are real. The pricing is structural. The companies that solve discovery and matching for these roles, operators, and compliance hires will be infrastructure for the agentic economy.

This is v0.3. It will be wrong in places. Tell me where.

Pull requests welcome at shipstacked.com/atlas.

About the author

Thomas Oxlee is the founder of shipstacked.com, the discovery and classification layer for the labor market of the agentic economy. He is currently embedded as the AI integration operator at a regulated EU business under AI Act exposure, where most of the field signal that informs this Atlas comes from. shipstacked.com matches AI-native specialists, agent operators, vertical specialists, and compliance practitioners to companies that need them — without CVs, without LinkedIn taxonomies, and without the assumptions of a labor regime that broke eighteen months ago.

shipstacked.com/atlas

v0.4 planned additions (folded in from later session decisions)

1. Domain practitioner with integrated AI as explicit, distinct supply category (lawyers, doctors, accountants, architects, financial advisors, etc. who have deeply integrated AI into their primary professional work). Pool: ~1.5-3M globally. Highest-LTV customer because OTHER practitioners in the same domain want to hire them. Distinct from Part VI vertical AI engineers — these are domain-first, AI-as-multiplier, not AI-first with domain knowledge.

2. ISCO-08 / SOC 2018 / O*NET crosswalk per role. Explicit mapping of each Atlas role to the closest existing official code, with gaps flagged. Makes the Atlas useful for visa applications, government contracts, EU AI Act conformity assessments where formal codes are required.

3. EU AI Act Annex III + ISO 42001 explicit mapping for Part III compliance roles. Cross-reference each Atlas C-cluster role to specific Article numbers and audit categories.

4. Expanded Levels.fyi-anchored compensation data per role per location per company. v0.3 has good data for FDE, Solutions Engineer, Deployment Strategist, Prompt and Context Engineer; v0.4 fills gaps systematically.

5. Acquisition-as-talent-supply pattern as named mechanism. Tomoro is the first public example (acquired by OpenAI May 11 2026 for 150 FDEs + Deployment Specialists). Expect more: Anthropic to acquire a similar UK/EU consultancy for its Blackstone venture within 6-12 months; Google DeepMind to acquire one; Mistral a European equivalent.

6. Three-layer venture structure as named pattern: frontier lab + direct-employed FDEs (Tomoro-flavored) + Big-3 consulting + Capgemini partner channel. The OpenAI Deployment Company explicitly has Bain & Company, Capgemini, McKinsey & Company as consulting/integration partners.

7. Named companies per vertical in Part VI expanded. Healthcare AI: Navina, Tempus, Butterfly, Commure, Foundation Health, Federato/Kalepa. Legal AI: Harvey, Cuckoo, Spellbook, EvenUp. And so on per vertical.

8. Expanded Part IV alignment research with specific lab program structures (Anthropic Fellows curriculum detail, MATS selection criteria, Redwood Residency structure, ARC research positions).

9. OpenAI Frontier / superapp framing implications for the Atlas. If OpenAI Frontier wins as the underlying intelligence layer governing company agents, the labor market shape changes: A1 shifts from custom integration to configuration/governance of Frontier inside the customer. D3 shifts toward configuring agent behavior within the Frontier framework. B-cluster shifts to monitoring at the Frontier governance layer. D5 gains importance because customers need lab-independent evaluation methodology.

10. "Hiding in plain sight" phenomenon named explicitly. Many practitioners doing genuine A1/A2/A4/A6 work are classified by their employer and by LinkedIn as something generic ("Senior Engineer," "Solutions Consultant," "Tech Lead"). The Atlas reframes their identity. The /claim form is the structural mechanism by which they self-classify into the routable supply pool.

— End of Atlas v0.3 —