Loading slide
Agents are not good at everything, but there is a clear shape of task they are genuinely excellent at. Learn the shape and you will know when to reach for one. Four features make a task agent-friendly, and the best tasks have all four.
The first is that the task breaks into clear steps. Research, then draft, then revise, then format. If you can name the steps, the agent can work through them one at a time instead of trying to leap the whole gap at once.
The second is that each step can be checked. Writing code is the model example: the agent writes a piece, runs it, sees whether the tests pass, and fixes what failed. The world answers back at every step, so the agent is not flying blind.
The third is that mistakes are recoverable. If a wrong turn at step four can simply be undone and retried, the loop's habit of building on its own steps stops being dangerous. Contrast that with sending an email or spending money, where step four cannot be taken back.
The fourth is that the task is mostly gathering and sorting information. Searching many sources, summarising long documents, pulling numbers from different systems: tedious for a human, and a natural fit for a loop that can fetch, read, and move on.
The thread tying all four together: an agent does best when it acts in a world that pushes back, giving it feedback and second chances at every step. The further a task drifts from that, fuzzy steps, no way to check, mistakes that stick, the more carefully you have to watch.