Playbook · Advanced · 5h
Multi-step agent with approvals: playbook
Implement an agent that plans multi-step work and pauses for a person before any side effect.
- Claude
- ChatGPT
- Any assistant
- agents
- playbook
Inside this playbook
- Multi-step tasks completed with every side effect approved by a person.
- Step 1: Separate planning from acting
- Step 2: Propose, do not act
- Step 3: Approve exactly what was proposed
- Step 4: Resume with the observation
- Step 5: Audit everything
- Steps
- 5
- Deliverables
- 3
The full playbook opens with a membership that covers AI Agent Implementation.
Sign in if you already subscribe.
Sign in