Where to find data on the AI agent job market

Open agentic roles

2,331

Tracked AI roles

3,515

Agentic share

66%

Companies hiring

193of 254 tracked

Remote share

9%

Snapshot

28 September 2026

What this is, and what it is not

Update cadence and freshness

How the index is built

1. 1. A curated watchlist, in two cohorts

   254 companies are tracked, classified the same way regardless of cohort: frontier(151 AI-native companies, seeded from the Forbes AI 50 and CB Insights AI 100) and enterprise (119 traditional enterprises, a curated Fortune 500 and ASX panel of visible AI deployers). 70 of the tracked companies are based in Australia or New Zealand, which is what makes the ANZ view possible. The watchlist is reviewed monthly; companies are added when they show verifiable agent engineering and removed when their careers feed goes dark.

2. 2. Postings fetched directly from each employer's applicant-tracking system

   Every Monday the pipeline reads each company's public ATS feed: Greenhouse, Lever, Ashby, Workday, Rippling, SmartRecruiters, Workable, Recruitee, Personio, BambooHR, Oracle and SuccessFactors, plus a browser-rendered tier for boards such as PageUp and iCIMS. It never reads LinkedIn, Indeed or any third-party job board, so there are no duplicates, no stale re-posts and no listings the employer has not published themselves. Boards that require a login or a rendered session with no public feed are marked unsupported rather than guessed at.

3. 3. An AI-specific title gate

   A posting enters the index only if its title is AI- or agent-specific: it must match one of the 8 category patterns below. Generic software-engineer, data-scientist, analyst, recruiter and sales titles are excluded, even at AI companies, unless the description names an agent framework, protocol or multi-agent architecture. Human "agent" roles (insurance, travel, real-estate, call-centre, procurement and similar) are excluded outright.

4. 4. What counts as agentic: one hard signal, or two distinct soft signals

   The full job description is screened after stripping company boilerplate from both ends (the "about us" intro and the benefits and equal-opportunity footer), because AI-forward employers name-drop their agent products in every posting, including non-technical ones. A posting is flagged agentic when it shows either:

   • one hard signal: a named agent framework (LangGraph, CrewAI, AutoGen or AG2, LlamaIndex, Semantic Kernel, OpenAI Agents SDK, Google ADK, Pydantic AI, Bedrock Agents, Agentforce, smolagents and others), an agent protocol (MCP, A2A, computer use), or an explicit multi-agent or orchestration architecture (supervisor or router agents, agent swarms, agentic workflows); or
   • two distinct soft signals from four categories: tooling (tool use, function calling), reasoning (planning, memory, ReAct loops, human-in-the-loop, task decomposition), outcomes (agent evals, observability, guardrails, agentic RAG, LangSmith or Langfuse), and the generic concept ("agentic", "AI agent", "autonomous agent").

   The concept category alone never qualifies: a generic "agentic AI" mention is not enough. Titles are also flagged on a precise pattern (for example "agent engineer", "forward deployed", "MCP"), never on a bare "agent" token. Every posting stores agentic_signals (the categories that fired) and agentic_source (title or description), so any classification can be audited from the download. In the current snapshot 78% of agentic roles rest on a hard signal. The classifier is deterministic (no language model), so the same input always yields the same answer and the rule can be published in full.

5. 5. Eight role categories

   Titles are matched against the categories in this order; the first match wins. Counts are all tracked roles in the current snapshot.

   Category	What it covers	Roles	Share
   Agent engineering	Building agents: agent engineer, agentic AI developer, multi-agent systems, MCP, forward-deployed engineer, copilot engineering.	2,280	65%
   AI/ML engineering	Foundational AI/ML: ML engineer, LLM engineer, applied AI, research engineer or scientist, prompt engineer, fine-tuning.	823	23%
   Agent ops & infra	Running agents: LLMOps, MLOps, AI platform and infrastructure, inference and model serving.	148	4%
   AI product	AI product manager, product lead or owner for AI, LLM or agent products.	107	3%
   Field & solutions	AI solutions engineer or architect, sales engineer for AI or agent products, applied AI for customers.	59	2%
   Evals & quality	Evaluation, AI quality, benchmarking.	49	1%
   AI security	AI, LLM or model security, red-teaming AI systems.	33	1%
   AI governance & risk	AI governance, responsible AI, model risk, AI policy, compliance, assurance and ethics.	16	0%

6. 6. Enrichment and publication

   Each posting is also tagged with seniority (from the title), a remote flag (from location and title), skills named in the description (MCP, LangGraph, RAG, evals, Python and others), cohort and location, plus first_seen andlast_seen. The result is written to data.json and data.csv, the summary is regenerated on the index, the analysis,the company tracker and the role guides, and on the first of the month areport is archived.

Known limitations

• It is a watchlist, not a census. Companies outside the 254 tracked are not counted. Totals describe agent hiring at the companies most likely to be doing it, not the whole economy.
• Precision over recall. The rule is strict, so the agentic count is a floor: some genuinely agentic roles with vague descriptions are left out, and the count is smaller than aggregator boards report.
• Openings, not hires. A posting counts while it is live on the employer's feed. The index measures advertised demand, not headcount.
• Feeds without descriptions are classified on title only. Those postings can never earn a description-based agentic flag, so cohorts on such feeds are under-counted rather than over-counted.
• Boilerplate is stripped heuristically. Where a company's fixed intro names its own agent product, a residual mention can still fire a signal. The per-posting signals make such cases findable in the download.

How it compares to other agentic AI job trackers

Three other sites publish counts of agentic AI roles. They are job boards built for job seekers, and they are good at that. They answer a different question from this index, and the differences below are why their numbers are larger. Details were checked on 16 September 2026; site claims are quoted as published.

Why the counts differ: aggregators pull from many job boards and tag roles by keyword, so the same posting can appear more than once and a generic data or software role with "agentic" in the boilerplate is counted. This index reads each employer's own feed once, screens the full description, and requires a hard signal or two independent soft signals. A daily refresh is useful for a job seeker; for a market statistic, a weekly snapshot with archived history and a published rule is what makes a number checkable a year later.

How it compares to labour-market research

Analyst and academic programmes cover AI and the labour market at economy scale. They answer the broad question (how AI exposure is changing work) on an annual or quarterly cycle. The Agentic AI Jobs Index is the live, company-level complement: specifically agentic roles, refreshed weekly, openly licensed.

Download & access

• /jobs/data.json — the full live dataset: every posting with company, title, category, cohort, seniority, remote flag, location, skills, first and last seen dates, and the agentic signals that fired.
• /jobs/data.csv — the same postings as CSV, for spreadsheets.
• /jobs/reports/latest.json — the most recent monthly report as JSON.
• /jobs/reports/index.json — the archive index of every monthly report.
• /jobs/report/ — the human-readable monthly report hub.
• /llms.txt and /llms-full.txt — plain-text summaries of this and every other index on the site, for language models and agents.

For the analysis built on this data, see the Agentic AI Jobs Index in depth, thecompany tracker, and every tracked role.