Learning path · Agents & Orchestration · 61
Planning and Reasoning
Decomposing goals into steps, choosing tools, and revising plans when observations contradict assumptions.
Why it matters
- Separates toy chatbots from task-completion systems.
- CoT and structured plans reduce derailment.
- Must be bounded to prevent runaway token spend.
Key ideas
- Task decomposition
- Replanning
- Explicit plans
Top resources
- 01PaperYao et al.
ReAct: Synergizing Reasoning and Acting in Language Models
Why this resource. Interleaving thoughts and actions instead of a single plan dump.
Covers in this concept
- ReAct
- planning
- 02ArticleAnthropic
Building effective agents
Why this resource. When a workflow with explicit steps beats an open-ended planner.
Covers in this concept
- workflows vs agents
Planning modules emit explicit step lists—search docs, compare options, draft email—for execution and audit. Replan when tools error or facts change. Expose plans to users for sensitive workflows so they can edit before execution. Cap replanning iterations; infinite "let me try another approach" loops burn budgets overnight. Persist plans as structured artifacts users can edit; free-form chat plans are hard to audit after incidents.
Updated 2026-08-09 · Full learning path