AGENT ATLASa field guide to agent workflows
Back to Single agent
Single agent

Plan & execute

Make the whole plan first, then carry it out step by step.

1 agentCore

How it works

  1. 1Planner step: produce an explicit ordered plan of sub-goals.
  2. 2Optionally show the plan for approval before any work begins.
  3. 3Executor step: complete each sub-goal in turn.
  4. 4Re-plan only if a step fails or reveals the plan is wrong.

Use it when

Multi-step tasks where committing to a plan up front reduces drift and lets you review the approach before any work happens.

Reach for something else when

Highly dynamic tasks where the plan will be wrong by step two — ReAct fits better.

Where you stay in the loop

You approve the plan before any work begins — that is the leverage point. Once executing, the agent runs; you re-enter only when it flags that the plan was wrong.

In the wild

"Migrate this codebase" → plan the file-by-file order → execute each, checking off the plan.

Hand this to your agent

Work in plan-then-execute mode.

First, output a numbered plan of sub-goals to achieve my objective — no work
yet, just the plan. Wait for my "go".

Then execute the plan one numbered step at a time, marking each done and
noting any deviation. If a step proves the plan wrong, stop and propose a
revised plan rather than improvising.

Objective: <...>

Replace the <…> placeholders, paste it into your agent, and it'll scaffold the workflow with you.