Agent planning

An agent that only picks the next tool can lose the global goal. An agent that writes a ten-step plan and never revises it will execute a plan the world already invalidated. Planning is the choice between those two, and the scratchpad that keeps the goal visible.

The big idea

Plan-then-execute. The model writes a multi-step plan first, then executes it. Good when the task is large and decomposable (research report, migration checklist). Weak when early observations invalidate the plan. You must allow re-planning.

Interleaved (ReAct-style). Plan only the next action given the latest observation. More adaptive. Easier to lose the global goal in a long trajectory.

Todo lists and scratchpads. Persist an explicit checklist in state, or expose a tool such as update_todo. Coding agents do this so the model can mark steps done without relying on attention over a 100k-token transcript. Compact the scratchpad when it grows. See context engineering.

A kitchen recipe is plan-then-execute. Cooking from whatever is in the fridge is interleaved. Real cooking does both: you sketch the meal, then adjust when the onions are already burnt.

Walking through state the prompt should not own

Long agents accumulate junk. Keep canonical state in a structure the runtime owns:

goal: "Debug flaky checkout test"
todos: [reproduce, find culprit, patch, run tests]
artifacts: [log_url, branch_name]
budget: {steps: 12 used 4, usd: 0.20 used 0.07}
last_error: null

Expose update_todo and attach_artifact as tools so the model writes here instead of hoping attention remembers. When you compact the transcript, reload goal, todos and recent errors so the agent does not forget the mission.

Common mistakes

• A rigid plan with no re-plan path. The first failed tool call makes the rest of the list fiction.
• No visible goal after compaction. The agent starts busy-work.
• A ten-step plan for "fix the typo in utils.ts." Match process weight to blast radius.

Check yourself

1. You need a human to approve work before any write tool runs. Which style? Plan-then-execute, so the plan is the approval artifact.
2. Why keep todos outside the transcript? Compaction and long context drop details. A structured checklist survives.

Key takeaways

• Plan first when you need approval or a large decomposition.
• Interleave when early observations will change the path.
• Many systems propose a plan, then ReAct inside each approved step.
• Keep goal, todos and artifacts in runtime state, not only in the prompt.

Further learning

How much autonomy to grant, and when a human must sit on the loop, is agent autonomy.