A Checklist for Delegating Work to an Agent
Agents are powerful, but delegating defenselessly causes misdelivery, info leakage, and runaway incidents. Here are the items to always confirm before delegating.
Pre-Delegation Checklist
- Are purpose and goal clear?: did you define deliverable form and completion conditions?
- Did you specify prohibitions?: stated not to do payment/sending/deletion/application
- Is privilege minimal?: did you narrow account/data/operation scope?
- Confidential boundary: is the info OK to input (confirm terms/internal policy)?
- Approval gate: did you insert human approval for external/irreversible actions?
- Verification means: how will you verify the deliverable's facts/figures?
- Rollback: is there a procedure to stop/revert on malfunction?
Operating Iron Rules
- Start small, observe behavior, then expand scope
- Keep logs (traceable later what was done)
- Default to "semi-automatic with human approval" over "fully automatic"
Chapter Summary
Automation/agents' value is freeing time. But irreversible operations are human, judgment is human, verification is human. Keep these three and you get safe, large efficiency gains.
