AI agent

An AI agent hands control of the sequence to the model. Instead of your code deciding which steps run, the model picks a tool, sees the result, decides what to do next, and repeats until it reaches the goal or a limit. That flexibility lets agents handle open-ended tasks where the path depends on what they discover along the way.

For product builders, the key question is whether you need one. Signals that you do: the goal is open-ended, the number of steps is unpredictable, the model must replan based on results, and the task runs long. Many features that get called agents are really workflows or a single tool-using call, and those are far easier to test, price, and support.

Agents change the engineering. Costs and latency vary widely per task, so you need step limits and budgets. Evals have to judge whole sequences of actions, not just a final output. And because the model chooses actions, permissions, sandboxing, and human approval for consequential steps become central.

Start with the simplest design that works and add autonomy only where it clearly pays. Agent design in depth is a subject of its own, beyond a single feature.