Skip to main content

​
Summary

Planning and reflection patterns add structure around an agent loop. Planning breaks a task into an explicit execution path. Reflection evaluates a draft or trajectory and decides what to improve.

​
Why It Matters

Pure step-by-step control is often not enough for longer or more structured tasks. The agent may need to see the whole shape of the work before acting, or it may need a deliberate review step after producing an initial answer. Planning and reflection are two ways to improve quality without pretending the first pass will always be good enough.

​
Mental Model

Planning patterns, such as the plan-then-execute style highlighted in the imported source material, separate work into two phases:
  • generate a task plan
  • execute against that plan
Reflection patterns add another loop:
  • produce an initial answer or artifact
  • critique it
  • refine it
These patterns are useful for different reasons.
  • planning reduces drift on multi-step structured tasks
  • reflection improves quality when first drafts are cheap but correctness or performance matters

​
Architecture Diagram

​
Tool Landscape

Planning works well when the task can be decomposed clearly:
  • structured research
  • multi-step analysis
  • code generation with explicit stages
Reflection works well when the system can meaningfully judge and improve its own output:
  • code quality and performance
  • report structure and completeness
  • task trajectories that need error correction
Both patterns benefit from lightweight memory because plans, critiques, and prior drafts need to be available without overwhelming the main context.

​
Tradeoffs

  • Planning improves coherence, but it can lock the system into a weak plan if replanning is impossible.
  • Reflection improves quality, but it increases latency and model cost.
  • Detailed plans help structured work, but they can become overhead for simple tasks.
  • Strong critique prompts can improve outputs, but they also create another place where prompt design can fail.
Useful defaults:
  • add planning when the task has more than one meaningful dependency chain
  • add reflection when quality matters enough to justify another pass
  • stop iterating when additional review no longer changes the decision

​
Citations

​
Reading Extensions

​
Update Log

  • 2026-04-21: Initial repo-native draft based on imported reference material and lab rewrite rules.