SKIP TO CONTENT

// ORIGINAL DATA · v1

The State of Agent Engineering Hiring, 2026

I pulled 119 agent/AI-engineer job postings from 32 employersusing only public ATS job-board APIs (Greenhouse, Lever, Ashby), coded each one against a fixed rubric, and asked a simple question: what actually separates "agent engineering" from a generic AI/ML job posting? This is a companion to the State of FDE study and a separate corpus from it. Everything here is computed from real fetched data. The dataset and methodology are free.

By Adam Boudjemaa · Snapshot 2026-07-24 · CC BY 4.0 · Extracted facts + link-backs only, no verbatim postings

Preliminary sample. 119postings is smaller than the FDE study's 519-posting bar. Read every percentage here as directional. Full caveats in Limitations.

Headline findings

81.5%

of agent/AI-engineer postings explicitly require an agent framework or agentic pattern

97 of 119 name LangChain, LangGraph, CrewAI, AutoGen, tool use, function calling, or multi-agent orchestration. That's the line between an agent-engineering job and a generic AI/ML one.

89.1%

require production coding

Whatever the title says, the real version of the role is still an engineering job.

35.3%

mention evals, red-teaming, or guardrails as part of the role

42 of 119. A third of postings already treat evaluation as a named responsibility, not an afterthought.

16%

of postings are based in France or elsewhere in the EU

19 of 119. Small enough that this is reported as one cut, not split further by country.

What these roles are called

"AI Engineer" is the single largest bucket at 42.9% (51 of 119postings), but a third of the corpus ("Other") uses a title outside the five target families entirely. The label hasn't settled yet.

AI Engineer42.9% (51)
Other34.5% (41)
Applied AI Engineer16% (19)
Agent Engineer4.2% (5)
Forward-Deployed AI Engineer2.5% (3)

What actually makes it "agent engineering"

Job titles drift faster than job content. This is the signal that matters: does the posting name an agent framework or agentic pattern at all?

Agent framework / agentic pattern named81.5%
Production coding required89.1%
Evals / red-teaming / guardrails mentioned35.3%
Autonomy / human-in-the-loop mentioned37.8%
INCLUSION GATE
Decide: Does this posting count as agent engineering?Decision tree: Does this posting count as agent engineering? If The title is in the agent family and the posting names an agent framework or agentic pattern, then Counted, with the agent signal set. If The title is in the agent family but no framework or pattern is named, then Counted, flagged as generic AI/ML. If The title is plain Machine Learning Engineer or Data Scientist, then Excluded by the title gate. If The source is LinkedIn, Indeed, Wellfound, YC or Google, then Never collected.DECIDEDoes this posting count as agentengineering?IF The title is in the agent familyand the posting names an agentframework or agentic patternCounted, with the agentsignal setLangChain, LangGraph, CrewAI,AutoGen, tool use, functioncalling, or multi-agentorchestration. 81.5% of the set.IF The title is in the agent familybut no framework or pattern isnamedCounted, flagged as genericAI/MLKept in the corpus so the 81.5% hasan honest denominator rather than ahand-picked one.IF The title is plain MachineLearning Engineer or Data ScientistExcluded by the title gateDeliberately narrower than MLhiring, which is a different andmuch larger market.IF The source is LinkedIn, Indeed,Wellfound, YC or GoogleNever collectedThose terms prohibit collection.Public Greenhouse, Lever and AshbyAPIs only.
How a posting was classified. The title alone never decided it: the agent-framework signal is what separates agent engineering from generic AI/ML work.

What they require

LLM STACK MENTIONS

RAG42.9%
OpenAI API22.7%
Anthropic / Claude21.8%
Vector DB (Pinecone, Weaviate, etc.)18.5%

TOP LANGUAGES

Python58%
TypeScript / JS29.4%
Go21.8%
SQL13.4%
C++4.2%
Java3.4%

58%

remote

31.9%

onsite

10.1%

hybrid

Experience floor where stated (n=82): median 5 years, p25 3, p75 6.

The France / EU cut

19 of 119 postings (16%) are based in France or elsewhere in the EU. The sample is too small to split further by country without implying precision it doesn't have, so it's reported as one cut.

Context: the demand signal is not just this dataset

This isn't Adam's data, and it isn't computed from the fact table above. It's cited, not asserted: Skills related to agentic AI (AI agents, agentic systems) grew from 0.06% of US job postings in 2024 to 0.23% in 2025. That is a year-over-year increase of more than 280%, about 90,000 postings. Lightcast, “Four Takeaways from the 2026 Stanford AI Index,” 2026.

Get the data

COLLECTION PIPELINE
Flow diagram: 6 stepsFlow diagram: 6 steps. Seed list, then Public ATS APIs only. Public ATS APIs only, then Title gate. Title gate, then Extract structured fields. Extract structured fields, then Temperature-0 keyword rubric. Temperature-0 keyword rubric, then 119-row fact table, 32 employers.1Seed list~170 mostly US, venture-backed,AI-native and infra companies, curatedby hand.2Public ATS APIs onlyGreenhouse, Lever and Ashby. Sourcesthat prohibit collection are nevertouched.3Title gateApplied AI Engineer, Agent Engineer,AI Engineer, Forward-Deployed AIEngineer, evals and AI-safetyspecialists.4Extract structured fieldsFacts and link-backs only. No verbatimposting prose is stored orrepublished.5Temperature-0 keyword rubricDerived signals (agent framework,evals, autonomy) codeddeterministically, auditable back totext spans.6119-row fact table, 32 employersPublished as JSON under CC BY 4.0.Never blended with the separate FDEcorpus at /state-of-fde.
Every figure on this page is computed from the 119-row fact table this pipeline produced. No step involves a judgement call that is not written down in the rubric.
DATASET SCHEMA
postings[]: one row per job posting
employerstring

Hiring company, plus its board slug and ATS platform.

posting_id / posting_urlstring

Stable ID and a link back to the original posting.

collected_atdate

Snapshot date. This is open roles at a point in time, not hires.

title_raw / terminologystring

The title as posted, and the normalised family it was mapped to.

location_raw / regionstring

Location as posted, and the coarse region bucket used for the EU cut.

remote_hybrid_onsiteenum

Work model, as stated by the posting.

yoe_minnumber | null

Minimum years of experience, only when the posting states one.

languages / cloud / llm_stack_mentionsstring[]

Technologies named in the posting.

agent_framework_mentionedboolean

The load-bearing signal: does it name an agent framework or agentic pattern?

evals_or_redteam_mentionedboolean

Evaluation, red-teaming or guardrails named as a responsibility.

autonomy_or_hitl_mentionedboolean

Autonomy or human-in-the-loop named.

production_coding_signalboolean

Whether the role actually requires shipping production code.

One row per posting in dataset-v1.json. Derived boolean signals are the coded fields; everything else is captured as collected.

Dataset (JSON)

All 119 rows, extracted fields + link-backs. CC BY 4.0.

Methodology

Sources, seed list, inclusion rules, coding rubric, limitations.

Limitations (read these)

  • Sampling frame: public Greenhouse/Lever/Ashby boards only, from a hand-curated seed list of about 170 mostly US, venture-backed, AI-native/infra companies. This is agent-engineering hiring among that population, not the whole market. It under-covers large enterprise on private ATS, APAC, and companies not on these three ATS platforms.
  • Preliminary sample size: 119 postings across 32 employers is smaller than the FDE study's 519/57 minimum-viable bar. Read every percentage here as directional, not precise, until a larger round runs.
  • Snapshot, not flow: point-in-time open roles are not hires. Trend claims need a second edition.
  • Title-gate precision: “AI Engineer” is a broad, fast-drifting title. The gate is deliberately narrower than plain “Machine Learning Engineer” to stay on the agent-engineering family this study targets, which means some genuinely agent-adjacent roles with unusual titles are excluded.
  • Derived fields (agent framework, evals, autonomy signals) are coded by a deterministic, temperature-0 keyword rubric, auditable back to text spans; not human-adjudicated for every row.
  • Never blended with the FDE study. The 519-posting dataset at /state-of-fde is a different corpus (forward-deployed-engineer title family); no row here is drawn from or merged with it.

Cite this

Boudjemaa, Adam. "State of Agent Engineering Hiring 2026," v1. adam-boudjemaa.com, 2026-07-24. https://adam-boudjemaa.com/state-of-agent-engineering. Data licensed CC BY 4.0.

▸ THE STATE OF FDE STUDYWHAT IS AN FDE?