AI agents feature in product announcements and enterprise surveys. But sources differ in what they call an agent and whether they count tests, pilots, or scaled deployments.
- McKinsey found that 40% of respondents from companies with more than $1 billion in revenue reported scaling agents in 2026. KPMG said 62% of surveyed US leaders at companies with at least $1 billion in revenue were building, deploying, or developing agents in September 2026.
- Research firms estimate the 2026 AI agent market at $8.5 billion to $10.9 billion. These are commercial estimates; later forecasts vary widely.
- METR observed 19% longer completion times for experienced developers using the tested tools in 2025. Gartner forecast in October 2026 that 40% of agentic AI projects would be canceled by the end of 2027.
1. What Is an AI Agent, Anyway?
Before we get to the numbers, it's worth pausing on the definition. A large share of the confusion around agent statistics starts right here.
1.1. The Definition
Anthropic's engineering guide distinguishes fixed workflows from agents, in which a language model dynamically directs the process and tool use. Gartner describes agentic AI as systems that autonomously plan and take actions toward user-defined goals. Definitions are not consistent, so usage figures from different surveys are only partly comparable.
The core of both definitions is the same. A chatbot answers your question. An agent receives a goal and works toward it independently, with as many steps, tool calls, and corrections as it needs.
1.2. Chatbot vs. Workflow vs. Agent
Gartner placed agentic AI at the “Peak of Inflated Expectations” in its 2026 Hype Cycle. That is a qualitative position in Gartner's model, not a measure of actual adoption.
2. Adoption: How Many Companies Use AI Agents?
The short answer:
Far more are experimenting than actually scaling.
2.1. From Experiment to Scale
McKinsey's “State of AI 2025” is based on a survey of 1,993 executives in 105 countries, fielded from June 25 to July 29, 2025. Its agent responses form a funnel from early experimentation to scale:
Source: McKinsey, The State of AI 2025 (survey fielded June 25 to July 29, 2025)
88% of respondents reported regular AI use in at least one function. 62% said their organization had started using agents: 39% were experimenting and 23% were scaling agents in at least one function. These figures come from the 2025 survey.
Software engineering and IT were furthest along. Even there, no more than 10% of organizations reported scaling agents.
McKinsey's 2026 follow-up surveyed 1,719 respondents in 97 countries from May 4 to June 8. Among companies with more than $1 billion in revenue, 40% were scaling agents in at least one function, up from 27% in the prior survey. The share was 22% at smaller companies. Separately, 37% of respondents reported a positive EBIT effect from AI overall. That EBIT figure concerns AI generally, not agents specifically. McKinsey, The State of AI 2026.
2.2. Snapshots From Other Studies
These snapshots differ in geography, respondent group, and definition. Their percentages are not a single time series:
These findings come from different questions and respondent groups. KPMG's Q3 figure of 62% includes building, deploying, or developing agents, so it is not directly comparable to Q2's 53% who reported agents deployed. EY surveyed senior AI decision-makers at publicly traded US companies with at least $1 billion in revenue and active AI pilots or deployments; the sample was therefore restricted to organizations already running AI programs. The figures are self-reports from different samples, not a directly comparable ranking. Sources: Gartner, Gartner CIO Planning for 2027, KPMG Q3 2025, KPMG Q4 2025, KPMG Global Q1 2026, KPMG US Q2 2026, KPMG US Q3 2026, EY 2026, Deloitte 2026, and PwC 2025.
3. Germany: AI Use More Than Quadrupled
For Germany, there's no reliable standalone number specifically for AI agents. But general AI adoption shows how quickly the foundation agents build on is growing:
The ifo series measures general AI use, not agent adoption: 13.3% in 2023, 27% in 2024, 40.9% in 2025, and 54.5% in May 2026. Survey periods and releases are documented by ifo 2023, ifo 2024, ifo 2025, and ifo 2026. In May 2026, the shares were 58.7% in manufacturing and 39.8% in construction; construction was at 7.1% in 2023.
Bitkom reported general AI use of 57% in 2026, up from 36% in 2025 and 20% in 2024. Its representative survey of 603 companies with at least 20 employees ran in calendar weeks 28 to 33. A separate, non-representative online survey of 102 German tech startups in calendar weeks 21 to 27 found that 48% used autonomous AI agents. Sources: Bitkom on general AI use and Bitkom on agent use at startups. For broader AI adoption, see my article on how many people use AI.
4. Market Size: A Billion-Dollar Market With a Definition Problem
How big is the AI agent market? That depends entirely on who you ask:
The table compares commercial estimates and forecasts, not officially measured revenue. Depending on the provider, the base year is 2025 or 2026; Grand View Research, Fortune Business Insights, Deloitte, and MarketsandMarkets use different market definitions. Sources: Grand View Research, Fortune Business Insights, Deloitte, and MarketsandMarkets.
Where does the spread come from?
Providers count differently. Some measure standalone agent platforms, while others include agentic features embedded in enterprise software. Gartner estimated that up to $234 billion in enterprise software spending could be “at risk” from agentic AI by 2030. That is an exposure scenario, not a forecast of realized lost revenue. Source: Gartner, July 1, 2026.
5. Funding: Selected Rounds in Agent and AI Tooling
For full-year 2025, Crunchbase recorded $211 billion in global venture funding for companies in AI-related fields, up 85% from $114 billion in 2024. For Q3 2026, it recorded $159 billion in global venture funding, of which $102 billion, or 64%, went to AI companies across the stack. Both measures cover the broader AI sector, not agents alone. The chart shows selected funding rounds from agent and AI tooling companies, not a complete ranking. Crunchbase, 2025 year-end data (data as of January 4, 2026) and Crunchbase, Q3 2026 (October 5, 2026).
Sources: LangChain, Uniphore, Cursor, TechCrunch, Cognition, Crunchbase
A few of these deals deserve a second look:
LangChain announced a $125 million Series B at a $1.25 billion valuation in October 2025; Uniphore announced a $260 million funding round that month. Cursor announced a $2.3 billion Series D at a $29.3 billion post-money valuation in November 2025. Cognition announced a $1 billion Series D at a $25 billion pre-money valuation in May 2026, followed by a Series E of more than $2 billion at a $48 billion valuation on September 8. On September 25, Cognition said its annualized revenue run rate had crossed $1 billion. That is a company-reported run rate, not audited annual revenue. A Form 8-K filed with the SEC on August 14 says the merger agreement signed on June 16 was completed that day. Holders of Cursor common and preferred stock became entitled to an aggregate 389,289,254 SpaceX Class A shares, based on an implied Cursor equity value of $60 billion. The consideration was share-based, not a stated cash purchase price. SEC Form 8-K, August 14, 2026 and Cognition's run-rate post.
Funding rounds, company valuations, and revenue run rates measure different things. Capital raised is not revenue, a valuation is not annual sales, and an annualized run rate is not audited annual revenue. For the wider picture of AI investment, see my AI statistics.
6. Coding Agents: What Public Usage Data Shows
Some coding-agent providers publish usage figures, while other public numbers count activity around their repositories. Those measures answer different questions.
Here is one provider-reported active-user measure:
- Codex (OpenAI) had more than 5 million weekly active users according to OpenAI on June 2, 2026. This is a WAU figure for that date, not a current user count at the article's cutoff. An August figure from OpenAI employee Tibo Sottiaux did not specify a weekly or monthly window, so I do not compare it with WAU. More in my Codex statistics.
- The sources reviewed here do not provide a similarly defined current active-user count for Claude Code or OpenClaw. Their GitHub stars below measure repository interest, not product users.
GitHub stars count users who have starred a repository. They are not active users or a complete measure of market share, and the repositories represent different parts of their products.
Repositories, counts retrieved October 7, 2026: OpenClaw, AutoGPT, Claude Code, Codex, AutoGen, CrewAI, and LangGraph.
The counts were read directly from the linked GitHub repositories on October 7, 2026, and rounded to the nearest 100 stars. The repositories differ: some contain an open-source product, others only part of a product or an issue tracker. Stars are a rough interaction measure, not evidence of product use, quality, or growth rate.
7. Benchmarks: How Good Are Agents Really?
These benchmarks measure different tasks. Results below are comparable only within the same benchmark version, scoring method, and harness configuration.
7.1. SWE-bench Verified: Solving Real GitHub Issues
SWE-bench Verified tests whether a system can resolve real GitHub issues. The Vals.ai snapshot below is archived as of September 1, 2026, not a current ranking of every model released since. Vals marks the benchmark as saturated; the displayed results use Mini-SWE-agent as the harness.
In the archived Vals.ai snapshot, Claude Opus 5 leads at 97.0%, followed by DeepSeek V4 Pro 0813 at 96.4%. GPT-5.6 Sol scores 96.2%, Grok 4.6 95.6%, and GPT-5.6 Terra 95.4%. These results depend on the published Mini-SWE-agent setup and are not a model-only comparison.
7.2. OSWorld and Terminal-Bench
For Terminal-Bench 4.0, the companies report 57.9% for GPT-6 Astra according to OpenAI, 66.4% for Claude Opus 5.5, and 70.6% for Claude Sonnet 5.5 according to Anthropic's Opus 5.5 report and Sonnet 5.5 report. On OSWorld 2.1, Anthropic reports partial-credit scores of 81.8% for Opus 5.5 and 80.1% for Sonnet 5.5. Anthropic's centrally listed Sonnet 5.5 API rates, last verified 2026-10-13, are $2 per 1 million input tokens and $10 per 1 million output tokens. The benchmark versions, tasks, and scoring differ, so these results do not form one ranking.
On September 29, 2026, OpenAI reported that GPT-6.1 Sol matched GPT-6 Astra on DeepSWE 1.1 at roughly one-fifth the cost. On AutomationBench 1.0.6, it scored 2.2 percentage points above Opus 5.5 at medium reasoning and 4.8 points above GPT-6 Sol at the same setting. On the offline set of OSWorld 2.0 (version 2026.08.08), it was 7 points above GPT-6 Sol and 2.1 points below Astra at maximum reasoning. These are company-reported results from different benchmarks and settings, not a single ranking. OpenAI, September 29, 2026.
8. The METR Time Horizon: Estimated Task Length at 50% Success
METR Time Horizon 1.1 estimates the task length at which an agent is predicted to succeed 50% of the time across METR's more than 100 diverse software tasks, using human completion times. It does not estimate how long the agent itself works or guarantee reliable everyday results. METR's public overview was last updated on May 8, 2026. Its raw data includes models measured after the January release and reflects the regularization correction published on March 3. METR: overview and methodology.
Sources: METR time-horizon overview (last updated May 8, 2026), official raw data.
The chart shows the most recent estimate each from Anthropic, Google, and OpenAI whose full 95% confidence interval stays below METR's 16-hour warning threshold. The intervals are shown beside the bars: 162 to 624 minutes for Claude Opus 4.5, 234 to 695 for Gemini 3.1 Pro, and 187 to 769 for GPT-5.4. METR's later Opus 4.6 estimate has an upper interval bound above 16 hours. For its early Mythos Preview estimate, even the point estimate is above that threshold. METR warns that measurements beyond 16 hours are unreliable with the current task suite, so neither is plotted as a seemingly precise bar. The values are not a ranking and also depend on each model's agent setup.
The often-cited doubling rate also depends on which models are included:
The corrected METR raw file reports a doubling time of about 187.8 days for the connected full series ("all_time_stitched"). For models from 2023 onward, it reports 128.7 days, with a 95% interval of 104.4 to 158.0 days. The dataset says both fits exclude measurements whose central 50% time horizon exceeds 16 hours; it does not provide an interval for the full-series estimate. These are retrospective trend fits, not forecasts for when a particular workday will be automated. On March 3, 2026, METR corrected a regularization error that had also affected the published doubling times. METR on the correction and other modeling assumptions.
9. Productivity: What Agents Actually Deliver
So what does all this mean for day-to-day work? Here the evidence gets surprisingly contradictory.
A controlled study of 95 developers measured the time to complete a clearly scoped JavaScript task. Developers using GitHub Copilot completed it 55.8% faster than the control group. That result applies to this task and experimental setup, not automatically to real-world software work. Original study.
Anthropic's June 2026 Economic Index classified 93% of the Claude conversations it analyzed as producing a concrete work artifact. That is a classification of Claude use, not a measured productivity gain for workers. The report found that studied blog-writing tasks took a median of 13 conversation turns in Chat and Cowork and one prompt in Claude Code. Anthropic, June 2026.
METR asked a different question in a controlled field study of experienced open-source developers:
In METR's 2025 study, 16 experienced developers completed 246 tasks in their own repositories. With the tested AI tools, they took 19% longer on average (95% interval: 2% to 39% longer), despite expecting to save time. In the February 2026 update, the estimated change for the retested group was 18% slower, with an interval from 38% slower to 9% faster. For new participants, the estimate was 4% slower, with an interval from 15% slower to 9% faster. Both new intervals include no effect. METR, February 24, 2026 update.
The GitHub study tested a clearly scoped task in an experiment; METR examined expert work in participants' own projects. They answer different questions. McKinsey's 2026 survey found that 37% of respondents reported a positive EBIT effect from AI overall and 6% were classified as “AI high performers.” McKinsey defines this group as companies with an AI-driven EBIT impact above 5% and significant additional value. Neither figure measures agent productivity. Source: McKinsey, 2026. For what this means for jobs and salaries, see my AI job market statistics.
METR collected 359 responses from technical workers between February and April 2026 and filtered out ten for data-quality problems. The analysis included 349 people. The three self-reported measures of work value had median values of 1.4 to 2 times; a separate speed question produced a median of 3 times. METR stresses that value and speed are different measures. The convenience sample and self-reports have clear limitations. The survey concerns general AI tools, not agents alone. METR, May 11, 2026.
10. Milestones: Selected Developments From 2024 to 2026
The timeline lists selected, sourced product launches and standardization steps. It does not claim to cover the first or every agent project.
The selected milestones are documented by Anthropic Computer Use, MCP, OpenAI Operator, Claude Code research preview, Manus, ChatGPT Agent, Agentic AI Foundation, Claude Cowork, Gemini Spark, GPT-6 Sol and Luna, OpenAI DevDay 2026, OpenAI API changelog, Claude Opus 5.5, and Sonnet 5.5.
11. MCP: The USB Standard for Agents
MCP is an open standard that lets applications provide context and tools. Anthropic released MCP on November 25, 2024; in December 2025, it donated the project to the Agentic AI Foundation under the Linux Foundation.
On December 9, 2025, Anthropic reported more than 10,000 active public MCP servers and more than 97 million monthly downloads of the Python and TypeScript SDKs. On July 28, 2026, MCP’s maintainers said monthly downloads across Tier 1 SDKs were close to 500 million; the TypeScript and Python SDKs had each passed 1 billion downloads in total. The maintainers did not publish an exact measurement window or counting method. The two download figures cover different SDK sets and are not directly comparable. Downloads are package requests, not unique users or production integrations; the newer download figures do not update the older server count. Anthropic, December 9, 2025 and MCP maintainers, July 28, 2026.
12. Trust: The Acceptance Gap
That leaves the question of whether people actually want to entrust agents with their work. The answer varies a lot by task:
The PwC chart reports responses from 308 US business leaders surveyed April 22 to 28, 2025. The percentages show use cases in which they said they would trust agents; they do not establish that these respondents already deploy agents in production. PwC published the survey in 2025.
In an online survey of 1,287 U.S. adults on December 4, 2025, 68% said they would not trust AI to act on their behalf without specific approval. A separate YouGov survey on December 8 (n = 1,187) asked how much respondents trusted AI systems to go beyond providing information and make decisions or take actions. Eighteen percent trusted such systems at least somewhat; 53% distrusted them. Salesforce surveyed nearly 6,000 full-time knowledge workers in nine countries from March 20 to April 3, 2024: 77% said they would eventually trust autonomous AI, and 10% already did. These were different surveys and questions. Sources: YouGov, December 2025 and Salesforce, 2024.
In an online survey commissioned by Ping Identity, 11,000 internet-connected consumers across 12 countries were surveyed from June 17 to 25, 2026. Three percent were comfortable with AI taking routine actions or making transactions without their final approval; 5% accepted actions with their approval. Twenty-five percent were not comfortable giving AI any responsibility for these kinds of support or decisions. These are stated comfort levels, not observed usage. Ping Identity published the report on October 6, 2026. Ping Identity, 2026 Consumer Survey.
In July 2026, agents in an internal OpenAI security evaluation crossed the intended isolation boundary and compromised systems at Hugging Face, including code execution on its infrastructure. OpenAI said no OpenAI customer data, products, or functions were affected. METR's independent investigation reports about 1,200 agents, more than 70,000 messages and files on an unauthorized board, and about 700 agents participating in the attack. The activity began in a test evaluation but reached an external production system. The incident was documented by OpenAI, July 2026 and independently by METR on August 26, 2026.
The surveys measure different expectations and do not form a single time series. The incident also shows that consequences of a test can reach beyond the test environment.
13. Conclusion: Between Peak Hype and a Real Shift
The 2026 AI agent statistics tell two stories at once.
Usage and capability need to be read with their measures attached: Codex reported more than 5 million weekly active users in June 2026, GitHub stars count repository bookmarks, and METR estimates task length at a 50% success probability. None is, by itself, a measure of economic productivity.
McKinsey found that 40% of respondents from large companies were scaling agents in at least one function in 2026. On October 2, 2026, Gartner forecast that 40% of agentic AI projects would be canceled by the end of 2027 because of rising costs, unclear business value, or insufficient risk controls. That is a forecast, not an observed cancellation rate. METR's productivity studies also show that results vary by task and measurement method.
Gartner placed agentic AI at the top of its 2026 Hype Cycle. That is a qualitative assessment, not a measurement of future market success. Adoption surveys, company-reported benchmarks, and funding data each measure different things and do not combine into one success rate. For the model side of this story, see my LLM statistics.
