Pattern · Agents
Plan, execute, reflect
A planner writes the steps, each step runs a short loop, and a check pass reviews the draft before the final answer.
Say it in the interview
For long tasks I would plan first, run each step as a short capped loop, then check the draft with code where possible and a model where not, allowing one revision.
When to use it
- Long research or multi-part tasks
- Pure ReAct drifts or loses the goal
How it works
- Planner writes steps → Step 1: short tool loop
- Step 1: short tool loop → Step 2: short tool loop
- Step 2: short tool loop → Draft
- Draft → Checks (totals add up, every claim sourced)
- Checks (totals add up, every claim sourced) → fail → One revision
- One revision → Final
- Checks (totals add up, every claim sourced) → pass → Final
Trade-offs
- Stays on track on long tasks
- More calls and latency than a single loop
Pitfalls interviewers probe
- Unlimited revision rounds
- Checks done only by the same model, never by code