An AI agent is a language model given tools and the right to take many steps in a row. It plans the work, runs commands, reads the result, finds mistakes and fixes them. One task, set in words, turns into hours of programming, data analysis or document review, and a person checks the outcome, not every step.
How it works
Inside is a simple loop. The model receives a task and a list of tools: a terminal, a browser, file search, an editor. It writes not an answer but an action — “run the tests”, “open the file”, “find the error in the logs”. A harness program carries it out and returns the result. The model reads what came out and chooses the next step. This repeats dozens and hundreds of times, until the task is solved or runs into a question for a person.
An agent’s strength is in checking. A programming agent does not take its own code on faith: it runs the tests, sees a red line, reads the error and fixes it. An analyst agent recalculates the total another way; a document agent checks the figures against the original source. On a long task the agent keeps notes — the plan, what is done, open questions — because the model’s memory is limited by its context window, while the notes stay on disk.
The agent gets exactly the permissions it needs: a sandbox instead of the production server, read access instead of write, a person’s confirmation before sending an email or making a payment. Every step uses tokens — the model rereads the task and the history — so a long task costs more than a short question. The price runs from $0.14 per million tokens for the open DeepSeek V4-Flash to $30 for the flagships.
Where it already works
- Programming: the agent writes code, runs tests and fixes bugs.
- Data analysis: export, cleaning, calculation and a report from one task in words.
- Document review: contracts, invoices, correspondence — cross-checks and summaries.
- Research: finding sources, comparing options and a final memo.
- Russia: agents on GigaChat, YandexGPT and open models on your own server.
What to consider
- Western cloud models are unavailable in Russia — GigaChat, YandexGPT and open models on your own server work.
- By METR’s measurements, the best models of 2026 succeed half the time on tasks as long as an expert’s working day — a person checks the outcome.
- Give the agent minimal permissions and a sandbox: it can delete or send just as fast as it can build.
- Count tokens: a long task uses several times more of them than a chat — compare rates before you start.