Use agents when the task needs actions, not just answers.

An agent is useful when AI must plan, use tools, inspect results and continue. The tradeoff is a larger blast radius, so permissions and verification matter.

Good agent tasks

• Repository changes with tests and reviewable diffs.
• Research loops that collect, compare and summarize sources.
• Recurring operational checks with clear exception rules.
• Browser or app workflows with bounded permissions.

Bad first agent tasks

• Irreversible production changes.
• High-stakes decisions with unclear approval points.
• Tasks with no measurable success condition.
• Workflows where the agent can silently broaden its own scope.