What Agents Can and Cannot Do Reliably

Agents are good at tasks that are loosely defined, take several steps, and have a clear way to check the result. They are weaker where a single wrong step is costly, where the goal is vague, or where the answer depends on facts the agent can't see.

A useful rule: an agent is reliable in proportion to how easily you can tell it got the answer right. Summarizing a document you can skim is easy to check. Booking a refund in a billing system you can't easily inspect is not. When checking is hard, add a human review step or use a fixed workflow instead.

Where Agents Fail in Practice

• They guess when they are missing information, and the guess sounds confident.
• They drift on long tasks and lose track of the original goal.
• They can take an action you didn't intend if your instructions leave a gap.
• They repeat a mistake if the tool keeps returning the same error.
• They can be steered by text inside the documents or pages they read.

None of this means agents aren't worth using. It means you design around these failure modes: clear goals, limited tools, checkpoints and review.

Exercise

• Pick a task you would like to hand to an agent.
• Write how you would check that it was done correctly in under a minute.
• If you can't write that, the task isn't ready for an agent. Shrink it until you can.
• Pass condition: a one-minute check written down for one task.