Plan-and-Execute Pattern

Plan-and-execute separates planning from doing. The agent drafts a plan, executes steps, and revises the plan when observations invalidate it.

When to use

• Multi-hop research
• Tasks with natural phases (gather → analyze → write)
• When you want an inspectable plan for HITL approval

Structure

1) Plan: [step1, step2, step3]
2) Execute step i → observation
3) If blocked: re-plan remaining steps
4) Finalize

Pros

• Better global coherence than pure greedy ReAct
• Plans are reviewable artifacts
• Easier to parallelize independent steps

Cons

• Bad plans waste work
• Re-planning adds latency
• Over-planning simple tasks

Production tips

• Keep plans short (5–9 steps)
• Attach success criteria per step
• Allow early stop when criteria met
• Use cheaper models for routine step execution

Variants

• Plan once (no replan) for stable environments
• Hierarchical planning for large projects
• Human-approved plans for risky domains