Planning: Reactive or Plan-and-Execute

Two ways to decide

A reactive agent picks the next step after each observation: flexible when the path is uncertain. A plan-and-execute agent writes a short plan first, then carries it out and revises when something fails: easier to review and to show the user, and cheaper because most steps need little extra reasoning. Many real agents mix both.

A plan as data

Keeping the plan as structured data lets code track progress and lets a human approve it before anything runs.

{
  "goal": "Reconcile last month's expenses",
  "steps": [
    { "id": 1, "do": "fetch_expenses(month=\"2026-09\")", "status": "done" },
    { "id": 2, "do": "match_receipts()",               "status": "doing" },
    { "id": 3, "do": "flag_mismatches()",              "status": "todo" },
    { "id": 4, "do": "write_report()",                 "status": "todo" }
  ]
}

Let people edit the plan

Showing the plan before running it lets a user fix a wrong assumption in seconds, which is much cheaper than discovering it after ten tool calls. Re-plan when a step fails instead of pushing on.

Quick check: What is a benefit of plan-and-execute?

• It always uses more tokens
• It needs no model
• It removes all errors
• The plan can be reviewed before it runs