Where agents work well

Agents are genuinely useful for a specific profile of task. Understanding that profile helps you know when to reach for one.

Decomposable tasks. If a task can be broken into clear steps — research, then draft, then revise, then format — an agent can work through those steps methodically.

Tasks where each step can be verified. Writing code is a good example: the agent writes it, runs it, sees if it passes tests, and fixes what doesn't. The environment gives feedback at each step.

Tasks with recoverable errors. If the agent makes a mistake at step four, it can backtrack and try again. This is different from tasks where an early error is permanent.

Information-gathering tasks. Searching across many sources, summarizing documents, pulling data from APIs — these are tedious for humans and easy to structure as agent loops.

The common thread: the agent benefits from acting in a world that pushes back. Not every task has that quality.