Planning

核心 · Key Idea

In one line: Planning means letting an agent decompose a big goal into actionable small steps first, then execute them step by step — the bridge between "can chat" and "can actually deliver."

What it is#

Complex tasks ("build me a SaaS landing page", "analyse this company's earnings") cannot be answered in one go. Planning breaks them down:

Goal: Build a SaaS landing page

Plan:
  1. Define the product positioning (talk to user)
  2. Draft an outline (Hero / Features / Pricing / FAQ)
  3. Write copy for each section
  4. Pick palette and fonts
  5. Emit the HTML

The agent then executes step by step, each step being a smaller, more tractable sub-problem.

Analogy#

打个比方 · Analogy

Planning is the construction crew's work order: "knock down walls first, then plumbing, then paint, then floors." Thinking the order through up front is faster and less error-prone than improvising on site.

Key concepts#

PlanPlan

An ordered list of steps, each an executable sub-task.

ReplanReplan

When the current plan turns out infeasible mid-execution, regenerate a new plan from fresh observations.

DecompositionDecomposition

Recursively break a big goal into a tree of sub-goals.

DAG / GraphDependency graph

Some steps in complex tasks can run in parallel; a graph reveals the dependencies.

Three mainstream patterns#

1. ReAct (think-and-act)#

Decide every step on the fly — planning emerges naturally. Simple but may zigzag with trial and error.

2. Plan & Execute#

Emit the full plan first, then batch-execute it. Saves Tokens, easier to observe, but mid-flight errors trigger replanning.

3. ReWOO / LLMCompiler (async / parallel)#

Compile the plan into a graph of parallel-executable nodes — many tool calls run simultaneously. Great for independent sub-tasks.

Practical notes#

Use only for complex tasks. Simple Q&A — don't plan, you'll just add overhead.
Make the plan a schema. Have the model emit JSON like [{step, action, args}] so a program can consume it cleanly.
Cap the step count. Prevent over-decomposition — 5–10 steps is usually enough; beyond that, switch to multi-agent.
Validate the plan once after generation. The cheapest Reflection trick: let the model look at its own plan and ask "did I miss anything?"
Replan on triggers, not every step. Replanning every step is expensive — trigger only on tool failure or unexpected observation.

Easy confusions#

Planning (ad hoc)

**Steps generated by the model at runtime.**
Flexible, handles novel tasks.

Workflow (fixed)

**Steps defined by humans at design time.**
Stable; guarantees critical steps are reached.