Planning agents — decompose, then execute

Vanilla ReAct decides one step at a time. For complex multi-step work this leads to drift, dead-ends, and rework.

Plan-then-execute:
1. Planner LLM produces a step list before any action.
2. Executor LLM (or the same one) runs each step with tools.
3. After each step, replan if needed.

Benefits:
- Plan visible upfront (auditable)
- Cheaper executor model possible
- Failures localised to a step

Use for: research tasks, coding tasks, anything with > 3 logical phases.

Use a smart model (gpt-4o, claude-sonnet) as planner — produces the plan once.

Use a cheap model (gpt-4o-mini, claude-haiku) as executor — runs each step. Cheap model is good enough for 'call this tool with these args' if the plan is clear.

Results:
- 5-10x cost reduction on long agentic flows
- Same or better outcomes if planning is good
- Faster iteration loop

This is the move that separates production agent systems from demos.

Planner returns a JSON list of steps. Executor loops over them, calling tools per step. Final synthesis combines results.

This 30-line skeleton handles 80% of multi-step agent flows. Wrap it with retries, cancelable steps, and tracing — that's a production planner.

Plans get stale. After every step, ask:
- Did the result change my model of the world?
- Are remaining steps still relevant?

If yes, replan. If no, continue.

A cheap classifier ('replan needed: yes/no?') after each step catches drift early. Without it, the agent rigidly follows the plan even when step 1's result invalidates it. Static plans don't survive contact with reality.

Day 73 done.

- Plan-then-execute > step-by-step
- Smart planner + cheap executor
- 30-line skeleton
- Replan after surprises

Tomorrow (Day 74): multi-agent systems — when one agent isn't enough.