Wiki
Core11 min read

Planning and decomposition

A step-by-step agent is short-sighted. Planning writes the subtasks and their dependencies first — and treats the plan as a hypothesis, not a script.

A purely reactive agent chooses each step by looking only at what has happened so far — which is short-sighted. Planning asks the model to write the shape of the work first: what the subtasks are, and which ones depend on which. The plan is not the answer; it is a map the agent follows, checks against reality, and revises.

Start here

Decomposition is the whole game. A goal that is too big to do in one step becomes tractable once you split it into pieces small enough to verify — and the dependencies between those pieces tell you what can run at once and which step is on the critical path.

One goal, decomposed — then run the independent parts at once

stage 1

scope the page

2u

parallel

write copy

3u

design layout

3u

stage 3

build page

4u

stage 4

test & fix

2u

stage 5

deploy

1u

makespan

12u

sequential 15u · parallel 12u · saved 3u

A plan makes dependencies explicit, which is what lets the independent branches run together — and what tells you which step is on the critical path. It is also a hypothesis: when a step returns something unexpected, the agent re-plans rather than grinding through a stale script. Durations are illustrative.

Toggle sequential and parallel execution on the same decomposition.

Plan-and-execute

The pattern is two phases:

  1. Plan — the model produces an ordered set of subtasks, often with dependencies, before touching a tool.
  2. Execute — the agent works the list, and when a step returns something surprising, it re-plans rather than forcing the old script.

This separates "what should happen" from "do it", which makes the plan inspectable — a human can read and correct it before any action is taken.

How to decompose

  • Least-to-most — solve the easiest sub-problem first, then use its answer to make the next one easier. Good when the sub-problems chain.
  • Hierarchical — break the goal into sub-goals, then sub-sub-goals, until each leaf is a single tool call or a short thought.
  • Dependency-first — identify what must happen before what; the ordering is often the hard part, not the individual steps.

Careful

A plan is a hypothesis about a task the agent has not done yet, so it can be wrong. Two failure modes: over-planning, where the agent spends its budget writing an elaborate plan for a task it could have just done, and premature commitment, where it grinds through a stale plan after the first step proved it wrong. The skill is a plan detailed enough to guide, loose enough to revise.

Check yourself

Eduspheria wiki · Agentic AI, Planning & reasoning

0 / 4 answered

  1. 1In plan-and-execute, what should the agent do when a step returns something surprising?
    Multiple choice
  2. 2Which decomposition solves the easiest sub-problem first and uses its answer to make the next one easier?
    Short answer
  3. 3A plan is a guarantee about how the task will unfold.
    True / false
  4. 4What is 'premature commitment'?
    Multiple choice

Next: the reasoning that happens inside each step.