What it is
An LLM agent is a language model placed in a loop where it can act: it thinks about what to do, calls a tool (a search, a calculator, a program), reads the result and decides the next step, until the task is done. ReAct showed the value of interleaving reasoning and actions; Toolformer showed that models can learn when to call tools.
What it is for
Tasks that need several steps and outside information: researching a question across sources, fixing a bug by running the tests, booking or filling in forms, analysing a spreadsheet. The more freedom you give an agent, the more useful and the more risky it is.
How to do it
- Define the goal and what "done" looks like in a way you can check.
- Choose the tools it needs, and no more: each tool has a clear name, a short description and inputs it expects (tools can come from MCP servers).
- Write instructions that say how to work: plan first, use the tools, check results, and ask you when unsure.
- Run it on an easy case and read every step it took, not only the final answer.
- Add limits: a maximum number of steps, a spending cap, and your approval before anything irreversible (sending, deleting, paying).
- Collect a handful of test tasks and re-run them after every change, because agents fail in new ways whenever the model or prompt changes.