Planner-executor pattern

• Tasks have branching paths and unclear best next step.
• You need transparency into what the agent intends to do.
• Retrying failed steps separately will save cost and time.

This pattern improves reliability and debuggability. PMs decide how detailed plans must be, how users approve them, and how to handle plan drift. It affects UX (show plan? allow edits?) and performance (extra turns vs. fewer failures).

Keep plans concise (3–7 steps) with clear preconditions and success checks. Cache the plan; retry failing steps before re-planning. In 2026, feed step outcomes into a small critic model to adjust future plans without user intervention.

A migration assistant plans schema export → risk checks → dry run → apply. Showing the plan to admins cut fear-based aborts by 40% and reduced failed runs by 18%, while adding only one extra turn (~400 ms).

• Allowing plans to be vague, giving no value for debugging.
• Re-planning on every failure, wasting time and tokens.
• Hiding the plan from users when the impact is high.