AI Agent Statistics 2026

How fast the agent layer is growing - search demand, hiring, protocol adoption and framework mindshare, from primary data.

Ethora Research · Last updated 2026-09-02 · Data snapshot: August 2026 · CC BY 4.0 - free to reuse with attribution

AI agents went from a research term to the centre of the conversational-AI category in about three years. The numbers below are primary measurements from The State of Conversational AI, Q3 2026 (Ethora Research) - search demand, open-job counts, package downloads, GitHub adoption and a capability audit of 18 systems - each linking to its full dataset. Data snapshot: August 2026; page reviewed monthly.

Key statistics

1. US search volume for "ai agents" grew about 700-fold - from ~60 searches a month in early 2023 to roughly 50,000 a month in 2026. Source: Search demand dataset
2. "AI agent" is a top-three US AI hiring category: roughly 61,000 open roles (Aug 2026), alongside generative AI (~81k) and AI engineer (~61k). Source: Jobs demand dataset
3. The Model Context Protocol (MCP), launched late 2024, hit 339 million monthly PyPI downloads ("mcp") and 198 million monthly npm downloads ("@modelcontextprotocol/sdk") by Aug 2026. Source: SDK downloads dataset
4. The modelcontextprotocol/servers repository reached ~89,561 GitHub stars less than a year after creation - the fastest riser in the tracked set. Source: GitHub adoption dataset
5. Agent-capable orchestration frameworks by GitHub stars: LangChain 144,246; Microsoft AutoGen 60,418; CrewAI 57,073; Flowise 55,362; LlamaIndex 51,637. Source: GitHub adoption dataset
6. In a capability audit of 18 notable systems, tool use and MCP support are near-universal, while proactivity (acting on schedules or triggers, not just replying) cleanly splits the field - it is where a chatbot becomes an agent. Source: Periodic Table dataset
7. RAG is a named requirement in roughly 19,000 open US roles; prompt engineering in ~20,000. Source: Jobs demand dataset
8. 24 of 28 agent- and chat-related buying terms now trigger a Google AI Overview - the category is discovered through AI answers. Source: Search demand dataset
9. Low-end agent implementation has commoditised: median advertised AI project on Freelancer.com is about $250 fixed-price, while skilled Upwork agent/LLM work runs a median ~$30/hour. Source: Freelance market dataset

Search demand: the 700x curve

Figure. Three-year US search-volume trend: "ai agents" leads the category's vocabulary. Volumes are third-party estimates.

"AI agents" is the fastest-growing term in the conversational-AI vocabulary: ~60 US searches a month in early 2023 to ~50,000 in 2026. For comparison, "mcp server" went from effectively zero before late 2024 to tens of thousands of monthly searches, while "rag" doubled and plateaued - retrieval became assumed infrastructure while agents became the story. Full 28-term dataset: search demand dataset.

Hiring: agents are already a top-three category

Figure. Open AI/LLM roles by country, absolute terms (Aug 2026).

In the US employer market (Aug 2026), the largest AI hiring categories are generative AI (~81,000 open roles), AI engineer (~61,000), and AI agent (~61,000) - followed by LLM (~35,000), prompt engineer (~20,000) and RAG (~19,000). The agent theme visible in developer interest and search demand now shows up directly in hiring. The US total is on the order of 300,000 open AI/LLM roles; per capita, Singapore leads at ~1,134 roles per million people. Full matrix: jobs demand dataset.

MCP: a protocol layer standardised in one year

The clearest structural change of 2025-2026 is the tools-and-protocol layer. The Model Context Protocol had no meaningful footprint before late 2024; by August 2026 the "mcp" package pulls ~339M monthly PyPI downloads, "@modelcontextprotocol/sdk" ~198M monthly npm downloads, and the modelcontextprotocol/servers repository sits at ~89,561 GitHub stars - top-of-registry adoption in under two years. In our 18-system capability audit, MCP or native tool use is supported nearly everywhere: a capability that did not exist before late 2024 is already near-universal.

MCP and agent-adjacent package downloads, last month (Aug 2026). Full table: SDK downloads dataset.

Agent frameworks and builders: developer mindshare

By GitHub stars (Aug 2026), the orchestration layer where agents are built is led by LangChain (144,246), Langflow (153,239) and Dify (152,436), with the dedicated multi-agent frameworks - Microsoft AutoGen (60,418), CrewAI (57,073), Flowise (55,362) and LlamaIndex (51,637) - forming the next tier. Stars measure mindshare rather than production usage; the download table above is the usage check. Full 34-project dataset: GitHub adoption dataset.

The capability frontier: proactivity and memory

We verified ten capabilities across 18 notable systems against their current documentation (Aug 2026). Tool use and MCP are now table stakes. Proactivity is the real differentiator: acting on a schedule or trigger, rather than only replying, cleanly splits the field - and it is precisely where a chatbot becomes an agent. Persistent memory is common in name but unfinished in practice (often an add-on, a paid tier, or a consumer feature not exposed to developers). No single system is strong on control, conversation and agency at once - that unresolved middle is the category's current frontier. Full matrix with per-cell sources: The Periodic Table of Conversational AI.

Frequently asked questions

How fast is AI agent adoption growing?

US search volume for "ai agents" grew about 700-fold between early 2023 and 2026 (to roughly 50,000 monthly searches), "AI agent" is a top-three US AI hiring category at ~61,000 open roles, and the MCP protocol layer reached 339 million monthly PyPI downloads within two years of launch.

How many AI agent jobs are there?

Roughly 61,000 open US roles mention AI agents (Aug 2026), making it a top-three AI hiring category alongside generative AI (~81,000) and AI engineer (~61,000).

What is MCP and how widely is it adopted?

The Model Context Protocol is an open standard (late 2024) for connecting AI models to tools and data. By August 2026 its Python package sees ~339M monthly downloads, its npm SDK ~198M, and its servers repository holds ~90,000 GitHub stars.

What separates an AI agent from a chatbot?

In a capability audit of 18 systems, proactivity - acting on schedules or triggers rather than only replying - is the cleanest dividing line, together with tool use and persistent memory. Tool use is now near-universal; proactivity and finished memory are not.

Numbers are point-in-time snapshots and interest proxies, not revenue or market share: stars and downloads measure developer attention and pulls, search volumes are third-party estimates, and job/talent counts are keyword- and LinkedIn-based with stated limits. See the methodology page for each metric's caveats.