Depending on which survey you read, somewhere between a tenth and a half of enterprises have AI agents in production. Surveys ask executives; executives are optimistic. Job postings are harder to inflate: a company that runs agents in production hires people to operate, secure and monitor them — and says so in writing. We read the agent passages of 36,357 live job postings and labeled what each role actually does with agents.
The answer is less hype and more plumbing. 3,172 companies are hiring to build agents for their own operations, 73% of them outside the tech industry. About one in five has at least one req for agents that are already live — and those live agents cluster in security and IT operations, not in the customer-facing showcases the demos are made of.
One in 76 job postings now mentions AI agents
In the last quarter of 2025, about 496 of every 100,000 job postings mentioned AI agents. By Q3 2026 it was 1,308 — a 2.6× rise in three quarters. Counted by company, the share of hiring companies with at least one agent-related req went from 1.6% to 3.3%: 18,492 companies in Q3 alone.
The fastest-moving piece of the stack is the plumbing. Mentions of MCP, the protocol that connects agents to tools and data, grew 3.1× as a share of hiring companies over the same period — faster than agents overall.
Past the demos? About one company in five
First, the noise. 35% of postings that mention agents describe no agent work at all — “agentic” sits in the nice-to-have list or the company blurb. Of the reqs that do describe real agent work, 11% describe agents that are already running: operating them, monitoring them, securing them, scaling them. The rest are building (most) or piloting.
By company, 628 of the 3,172 in-house builders — 20% — have at least one req for live agents; 298 are only exploring. Tech leads, but not by much: tech 24%, finance 21%, everyone else 17%, health & pharma 16%. Live agents also skew large — the share rises with company size — which fits Gartner’s warning that over 40% of agentic projects will be scrapped by the end of 2027: the projects that survive are the ones with operating budgets behind them.
Agents go live in security and IT first
Sort business functions by how often the agents are already running, and the back office wins. Security leads: 32% of companies building security agents have a req for live ones — SOC triage, detection, and, increasingly, securing the agents themselves. Uber’s Staff Security Engineer req: “perform security design reviews and threat modeling for critical services and AI agents”. IT operations is next at 25%: TikTok’s EMEA IT Operations Lead req: “Manage, maintain, and continuously improve agentic AI and automation workflows that handle baseline IT operations, including ticket triage, monitoring responses, routine remediation”.
Compliance and legal (24%) follow — New York Life’s VP, Head of Responsible AI Engineering req: “the governance itself will increasingly be delivered through AI agents operating across the software development lifecycle”. Customer service, the most-marketed agent use case of all, is built in-house by only 265 companies — fewer than finance or compliance — consistent with it being bought from vendors rather than built. And the long tail is real: T-Mobile’s Technology Sourcing Manager req: “Deploy and orchestrate AI agents across the full sourcing lifecycle”; Tesla’s Software Engineer, Data Analytics Agents & Platform req: “production LLM agents for conversational, tool-augmented analytics on complex, large-scale, governed data—not dashboards alone”.
The biggest single category isn’t a function at all. 1,184 companies — 37% of in-house builders — are building a shared agent platform before committing to use cases. JPMorganChase’s Lead Software Engineer - ML Engineer for Agent Platform req: “agent execution and sandboxing (micro-VMs), A2A and MCP integrations, the memory layer”. Once it’s live, it becomes an operations job: Southwest Airlines’ Tech Lead Software Engineer – AI Ops/Support req: “operate, monitor, and support Southwest's AI and agentic capabilities around the clock”.
| Function | Companies | With live agents | Live share |
|---|---|---|---|
| Security (SOC, securing agents) | 496 | 161 | 32% |
| IT operations & service desk | 657 | 161 | 25% |
| Compliance, risk & legal | 267 | 63 | 24% |
| Customer service | 265 | 56 | 21% |
| Shared agent platform | 1,184 | 230 | 19% |
| Finance & accounting | 289 | 42 | 15% |
| Software engineering | 914 | 124 | 14% |
| HR & recruiting | 83 | 12 | 14% |
| Insurance claims & underwriting | 58 | 8 | 14% |
| Sales & marketing | 433 | 52 | 12% |
| Clinical & patient workflows | 78 | 9 | 12% |
| Documents & knowledge | 465 | 51 | 11% |
| Data & analytics | 901 | 92 | 10% |
| Research & science | 203 | 15 | 7% |
| Supply chain & operations | 127 | 9 | 7% |
The stack: Microsoft is winning the platform war, MCP the protocol war
Platforms. Outside the vendors’ own hiring, Copilot Studio is named by 2.3× as many companies as Agentforce (1,585 vs 675 in Q3). Among in-house builders, 89% of Copilot Studio companies are outside tech — it is the low-code route for banks, manufacturers and utilities.
Frameworks. LangChain still has the most companies, but its share of hiring has been falling since Q1 while LangGraph keeps climbing: 0.37 LangGraph companies per LangChain company in Q4 2025, 0.68 by Q3 2026. Builders are moving from chains to graphs — explicit state, branching and retries, the things production agents need.
Protocols. MCP is named by 9.6× as many companies as Google’s A2A (5,820 vs 604), and companies that name it have one of the highest live-agent rates in the table (35% vs 20% overall). Connecting agents to real systems is what production looks like.
| Technology | Companies | Outside tech | With live agents |
|---|---|---|---|
| LangChain | 347 | 71% | 24% |
| LangGraph | 272 | 69% | 28% |
| CrewAI | 110 | 70% | 20% |
| AutoGen | 103 | 76% | 18% |
| LlamaIndex | 94 | 66% | 27% |
| Semantic Kernel | 61 | 72% | 16% |
| MCP | 253 | 67% | 35% |
| Copilot Studio | 282 | 89% | 21% |
| Agentforce | 77 | 65% | 22% |
| n8n | 136 | 65% | 25% |
| Power Automate | 201 | 92% | 18% |
| UiPath | 46 | 85% | 17% |
| Amazon Bedrock | 161 | 75% | 30% |
| Azure OpenAI | 114 | 82% | 21% |
| Azure AI Foundry | 86 | 86% | 19% |
| Vertex AI | 94 | 71% | 29% |
| Langfuse | 49 | 73% | 41% |
| LangSmith | 28 | 68% | 46% |
| Pinecone | 41 | 71% | 34% |
| Temporal | 25 | 48% | 32% |
The observability gap: live agents, few instruments
If agents are going live, someone should be measuring them. Among in-house builders with live agents, 40% mention evals and 45% mention guardrails, against 19% and 25% of companies still building. But only 4.8% name a dedicated agent observability or evaluation tool (LangSmith, Langfuse, Arize, Ragas and peers). The leaders are building the instruments themselves — PepsiCo’s Deputy Director - AI Solution and Platforms req: “production-grade instrumentation, telemetry pipelines, dashboards, quality gates, and safety signals across agentic AI systems”; Kraft Heinz’s Group Lead, AI Governance req: “prompt and response tracing, tool-call visibility, evaluation datasets”.
That is the gap behind the cancellation forecasts: most companies putting agents in front of employees and customers have not yet hired for the job of knowing whether they work.
| Excluded | Why it’s out |
|---|---|
| “Agentic” as a buzzword | 12,664 reqs (35%) mention agents only as a nice-to-have skill or in company boilerplate. Counting them would inflate the universe by two-thirds with no agent work behind it. |
| Human and software “agents” | Insurance, travel and real-estate agents; monitoring agents (Dynatrace OneAgent, SQL Server Agent, Elastic Agent); open-policy agents. Dropped by the keyword screen and the classifier. |
| AI-trainer gigs | “Train and evaluate AI agents” data work for AI labs and data vendors — screened by title and by the classifier. It builds someone else’s model, not the hiring company’s agents. |
| Consultancies, IT services and agencies | 13,830 postings: they build agents for clients, so their reqs describe someone else’s stack. |
| Vendors’ own hiring | 2,390 companies whose agents are the product are counted apart from in-house builders. Vendors’ own postings (Salesforce, Microsoft, LangChain, Anthropic, Anysphere and others) are excluded from the platform and tool comparisons. |
| Coding-agent users | 3,276 reqs where the role uses Claude Code, Cursor or Copilot but builds no agents — reported in the sidebar, not in the builder counts. |
628 companies are running agents in production right now
The free CSV lists the 50 largest in-house builders with live agents — the req, the business function, and the stack their postings name. The full list, by stack and use case and refreshed quarterly, is the product: if you sell agent observability, evaluation, security or identity, MCP gateways or agent platforms, these are the companies that just took on the problem you solve. Custom pull against your account list is free.
How this was measured
Echoloc continuously indexes public job postings; at analysis time: 24,695,997 postings, 3,358,623 active, across 1,074,166 companies.
- Screen: active postings whose title or description mentions AI agents (“agentic”, “AI agent”, “multi-agent”, LLM or autonomous agents) or names an agent-specific technology (LangGraph, CrewAI, AutoGen, Semantic Kernel, MCP, Agentforce, Copilot Studio and others) — 59,712 postings.
- Exclusions: consulting, IT-services, staffing, outsourcing and agency industries (15 categories) plus named consultancies; AI-trainer gig titles. Function screen: technical, data, product, security and IT roles, or titles naming AI or automation — 36,357 reqs.
- Labeling: the agent passages of every req were read by a language model (Claude Haiku 5.5) against a fixed rubric — stage (live, building, piloting, tool use, buzzword, noise), who the agents serve (own operations, own product, clients) and up to three business functions. A hand check of the “live” label found most cases unambiguous; the borderline ones are mainly security roles hired to secure agent systems, so read “live” shares as an upper bound.
- In-house builders: companies with at least one req building agents for their own operations, where own-operations reqs are not outnumbered by product-facing ones. Companies selling agents (2,390), building for clients (335) or without a clear signal (1,465) are kept apart.
- Momentum uses mentions in all postings published per month (not just active ones), as a share of postings or of hiring companies; months before Jan 2026 enter via still-open postings. Company comparisons for platforms and coding tools exclude the vendor’s own postings.
- Counts are companies, not postings, except where postings or reqs are stated.
Questions, or a number you’d like re-cut? [email protected].