AI agents — systems that take multi-step actions toward a goal rather than answering questions — became the dominant startup category of 2025, with agent-related companies absorbing a large share of all AI venture funding during the year, per Reuters and industry survey reporting. The pattern through late 2025 is consistent across the sector: demos impress, pilots proliferate, and paid deployments concentrate where errors are cheap and success is checkable. This is an analysis of the documented market, not a prediction, and not investment advice.
What actually counts as an agent?
A useful working definition on the record: an AI system that plans, uses tools — browsing, APIs, code execution — and acts across multiple steps with limited supervision. Chatbots that answer are not agents; systems that book, file, code, or purchase are. The 2025 model releases that pushed the category were the major labs' reasoning and tool-use models, which made long action chains technically feasible for the first time. Feasible is not the same as reliable, and reliability is the line every agent startup is now priced against.
Where are agents actually earning revenue?
The documented revenue cases cluster in four settings. Coding: developer tools with agent modes — Cursor's commercial traction being the most cited, alongside GitHub Copilot's enterprise expansion — where the user checks the output and errors cost minutes. Customer support: deflection of routine tickets, where outcomes are measurable and failures hand off to humans. Back-office document processing: regulated workflows with human sign-off, from invoice matching to insurance claims triage. And sales prospecting research, where volume tolerates error rates. The common thread: the buyer can verify the work cheaply. Categories where verification is expensive or errors are catastrophic — medical, legal drafting, autonomous trading — remain dominated by copilots and pilots, not deployments.
What is the error-rate problem in numbers?
The scaling problem is arithmetic. A task chain with ten steps at 95 percent per-step reliability completes successfully about 59 percent of the time; at 99 percent per-step, about 90 percent. Enterprise buyers quoted across 2025 industry surveys commonly put autonomous workflows in production only where per-step reliability approached the high 90s or where a human remains in the loop. This is why the measured benchmarks that matter commercially are long-horizon task completion rates — the metric the major labs began publishing with their 2025 reasoning-model releases — rather than single-question accuracy. Startups whose pitch survives that metric shift are the ones raising follow-ons; the rest are quietly repositioning as copilots.
How is the competitive structure settling?
Two layers. Model providers kept absorbing capabilities — every major 2025 model release included stronger tool use, compressing thin wrappers whose only product was orchestration around a single model. What survived at the application layer, on the documented funding record: companies with proprietary workflow data, deep integrations into systems of record, or human-in-the-loop operations that deliver guaranteed outcomes — plus agent infrastructure, the unglamorous layer of observability, evaluation, and permissioning that emerged as its own funded category in 2025. The agent stack is beginning to look like the cloud stack: infrastructure margins, application lock-in, and a graveyard in the middle.
What does the buyer side show?
Enterprise procurement moved from experimentation budgets to line-item evaluation during 2025, with the recurring pattern reported by vendors and buyers alike: a paid pilot of three to six months, judged on completion rates against a human baseline, then either expansion or quiet cancellation. The startups winning expansions share documented traits: they price on outcomes or seats rather than tokens, they deploy inside the customer's security perimeter, and they show the audit trail — increasingly a compliance requirement, since the EU AI Act's August 2025 obligations for general-purpose AI include documentation that flows down to application builders.
What should founders in the category watch?
Three measurable things rather than sentiment. Long-horizon completion rates on public benchmarks, which set the ceiling for what orchestration can sell. The labs' tool-use releases, each of which obsoletes a layer of the middle. And the ratio of pilot revenue to deployed revenue in any agent startup's numbers — the single best proxy for whether the product works unsupervised, and the number most often hidden by the phrase 'customer count.' The demo wave is over; the completion-rate wave is the market.
The honest summary of the record: agents work where verification is cheap, struggle where it is not, and are being priced accordingly — finally, by metrics rather than by video.
For more context, read Robotics Funding Wave: Figure and the $39 Billion Re-Rating.
For more context, read eu ai act startups.
For more context, read Datacenter Energy: The Wall AI Startups Are Heading Toward.