StoryboardAI Engineering

Planning

The agent writes itself a to-do list, works through it, and rewrites it when reality disagrees.

What is planning?

On a job with many steps it is easy to lose track of what has been done and what is left. So agents write a short plan first: a to-do list in plain words.

The list is not a script. After each step the agent looks at what happened and updates it: ticking items off, adding ones it did not foresee, dropping ones that turned out not to matter.

A visible plan also helps the person watching. It shows what the agent thinks the job is, which is the quickest way to notice it has misunderstood.

Plan first, then act.Step 1: Explore without changing anything. Step 2: Write the steps down. Step 3: A person approves the plan. Step 4: Work through it, ticking steps.1Explore withoutchanging anything2Write the stepsdown3A person approvesthe plan4Work through it,ticking steps
Plan first, then act.

Why does planning matter?

An agent that chooses each step only from the last result can wander: fixing what it just saw and losing sight of the goal. On long jobs, the worst mistakes are made in the first minute, by starting in the wrong place. A plan, written down and kept in view, keeps a long job pointed at its purpose.

How does planning work?

  1. Before acting, the model is asked to break the task into steps.
  2. The plan is stored by the harness, outside the conversation.
  3. It is shown to the model again on every round, with finished steps marked.
  4. The model may revise it as it learns more, and the harness records the change.
  5. For risky work, a person reads and approves the plan before any action is taken.

When should you use planning?

Use it when

  • A job has many steps or will run for a long time
  • Mistakes are expensive to undo
  • Several agents need to divide the work

Think twice when

  • The task is one or two steps
  • Nothing is known yet. Explore first, then plan

Where does planning sit in a real system?

The plan lives in the harness as a small list. Because the harness re-shows it every round, it survives even when older parts of the conversation are summarised away.

What goes wrong with planning?

  • Making a plan once and never looking at it again.
  • Treating the plan as fixed when new facts show it is wrong.
  • Letting the model tick off a step that code could verify.

What does planning look like in practice?

Asked to change all prices from dollars to cents, an agent first reads the code without changing anything and writes a plan: the price type, then cart and checkout, then coupons, then emails, then all tests. A person corrects one step. Only then does it begin.

Watch it happen

These stories show planning at work, one step at a time.