What an AI agent actually does all day

A chatbot answers. An agent acts. The technical difference is a loop: the model produces a step, something executes it, the result comes back into the context, and the model produces the next step. Give it a browser, a terminal, a database connection and a calendar, and it can do multi-step work without a person between each step.

The capability is real and narrower than the marketing. Understanding where the line falls saves a great deal of money.

Where agents are genuinely good

Software work with a fast feedback signal. Writing code, running tests, reading the failure, fixing it. This is the best case because the environment tells the agent whether it was wrong within seconds, and that check is objective. Most of the measurable productivity evidence comes from here.

Research with citations. Searching, reading, extracting and assembling a summary with links. The work is verifiable by a human in less time than doing it from scratch.

Structured drudgery. Cleaning inconsistent data, reformatting documents, filling forms from an inbox, triaging support tickets by category. High volume, low stakes per item, easy to spot-check.

Where they break

Any task with a long chain and no intermediate check. Errors compound: an agent that is 95 percent reliable per step is 60 percent reliable over ten steps and 13 percent over forty. This single piece of arithmetic explains most disappointing pilot results.

Tasks where the environment cannot say no are the dangerous ones. An agent that writes a file, sends a message or moves money has no test suite to fail. That is why the practical design pattern that survived contact with production is narrow tools, explicit confirmation for anything irreversible, and a log of every action.

And tasks requiring context the model does not have. An agent does not know that this customer is always late, that the finance team hates that format, or that the second database is deprecated. Unless it is told, it will act plausibly and wrongly.

The cost model nobody budgets for

Agents are expensive in a way chat is not. A single agent task can consume hundreds of thousands of tokens, because every tool result comes back into the context and the whole context is reprocessed on each step. Cost therefore scales with the square of the conversation length in the worst case, not linearly.

Three levers matter. Caching the stable part of the prompt cuts input cost substantially and is the first thing to implement. Model routing, sending easy steps to a small model and hard ones to a frontier model, is the second. Step limits are the third, and the most neglected: an agent with no budget will happily spend forty steps on a task that was not going to work.

Our AI inference cost calculator is built for exactly this arithmetic, and the slider that surprises most people is cache hit rate.

The measurement problem

Benchmarks for agents test tasks with a clear success condition, which biases them towards the cases agents are already good at. Real work is mostly ambiguous. If you are evaluating agents in your own organisation, measure the thing you actually care about: tasks completed without human correction, and time spent reviewing output. The second number is the one that quietly eats the savings.

What to watch

Watch for published reliability over long horizons rather than single-step benchmarks. Watch whether standard tool interfaces settle, because that determines whether agents become composable or stay bespoke. And watch the audit story: agents will not be trusted with anything consequential until what they did is as reviewable as what a person did.