Planning and decomposition as control structures for agents

Core idea

A plan is not proof that an agent knows what it is doing.

A good plan is a control structure. It helps the agent decide what to do next, what to check, when to stop, and when to ask a human.

Too little planning creates chaos. Too much planning creates rigidity. The goal is controlled adaptation.

The loop

Treat a plan like a hiking route, not a script. You want a map, but you still need to notice weather, blocked paths, fatigue, and wrong turns.

goal
  ↓
break into small steps
  ↓
do the next step
  ↓
check reality
  ↓
keep, revise, stop, or ask

The plan controls the work. Feedback controls the plan.

What decomposition does

Decomposition splits a goal into smaller work units.

Small units are easier to:

• inspect
• test
• pause
• recover from
• explain
• hand to a human for review

"Build the lesson" is hard to control.

A better version:

1. define the learner outcome
2. draft the core explanation
3. add one example
4. add a reader-facing practice step
5. check public safety
6. verify the route builds
7. ask for review before publishing

Each step has a purpose and a check.

The four controls

A useful agent plan answers four questions:

1. Order: what happens first, next, and last?
2. Scope: what is inside the task, and what is out?
3. Verification: how will the agent know a step worked?
4. Escalation: when should it stop and ask a human?

That makes planning part of reliability, not just organization.

Failure modes

Too little structure:

• starts before understanding the goal
• calls tools without knowing what result is needed
• fixes symptoms instead of causes
• produces partial artifacts with no finish line

Too much structure:

• continues after a prerequisite fails
• ignores test results because the plan says to move on
• treats a checklist as judgment
• avoids asking the human at the risky moment

Good planning sits between drift and autopilot.

Human checkpoints

Human review belongs inside the plan.

Agents can draft, test, and revise low-risk work. Higher-risk actions need explicit review before they happen:

• publishing
• sending messages
• spending money
• deleting data
• changing production systems
• presenting uncertain claims as facts

The point is not to slow everything down. The point is to put judgment at the boundary where consequences increase.

Builder checklist

Before letting an agent run a plan, ask:

1. Is the goal specific?
2. Are the steps small enough to verify?
3. Is there a stop condition?
4. Are risky actions gated by approval?
5. Can the agent revise the plan when evidence changes?
6. Is the final artifact observable?

If not, the plan may create motion without control.

Beginner checkpoint

Answer in your own words:

1. Why is a plan not proof?
2. How does decomposition improve verification?
3. What is one sign of a chaotic agent?
4. What is one sign of a rigid agent?
5. Where should human approval appear?

Turtleand take

Planning is the steering layer between intention and action.

Useful agents do not worship plans. They use plans to keep work bounded, observable, reversible, and reviewable while still updating from reality.