You finish with a mission brief, permission matrix, approved plan and the record of one small, reversible run.
Playbook: governed agentic delegation
Nine practices for designing, running and reviewing a workflow without confusing autonomy with absent control.
- Practice
- Choose a case, design the contract, limit tools, approve the plan, observe execution and decide whether to keep, adjust or stop.
- Transfer
- Expand autonomy only after the pilot passes; otherwise adjust or stop.
Delegate a process without giving up your judgment.
Giving an AI a vague goal can accelerate the wrong outcome. Useful autonomy begins by stating what it may decide, what it must escalate and how success is checked.
Choose a case, design the contract, limit tools, approve the plan, observe execution and decide whether to keep, adjust or stop.
Keep the brief, approved plan, exceptions, sources used, result and final decision.
Full index13
Starting point
Design a bounded first agentic workflow: clear objective, minimum permissions, plan approval and evidence before expanding autonomy.
Nine practices for designing, running and reviewing a workflow without confusing autonomy with absent control.
- Challenge
Giving an AI a vague goal can accelerate the wrong outcome. Useful autonomy begins by stating what it may decide, what it must escalate and how success is checked.
- Practice
Choose a case, design the contract, limit tools, approve the plan, observe execution and decide whether to keep, adjust or stop.
- Evidence
Keep the brief, approved plan, exceptions, sources used, result and final decision.
Choose the case
Principle: Start where failure is visible and recoverable.
A stable case tests the contract without mixing novelty, sensitive authority and hidden dependencies.
Go deeperHow to apply it
How to apply it
- Compare three candidates.
- Reject external effects and high sensitivity.
- Record your non-delegable boundary.
Checklist
- Repeatable, bounded case.
- Personal boundary written.
- Decision rule
- If you cannot observe and recover, it is not the first pilot.
- Example
- Synthetic example: draft from private notes, without sending.
- Antipattern
- Choosing the most complex process to signal ambition.
- Expected evidence
- Suitability matrix and rejection reasons.
- Limit
- Suitability applies only to the assessed case and scope.
Define owner and result
Principle: Human accountability must be nameable.
The agent produces or proposes; a person answers for the objective and a separate review may test the result.
Go deeperHow to apply it
How to apply it
- Name the accountable owner.
- Name a reviewer when needed.
- Retain your personal approval.
Checklist
- Explicit owner.
- Final approval reserved.
- Decision rule
- No agent is owner, approver or source of authority.
- Example
- Synthetic example: you accept the draft and record why.
- Antipattern
- Using “the team” as owner without a name or gate.
- Expected evidence
- Accountability card and verdicts.
- Limit
- Human review cannot repair a vague contract by itself.
Prepare context and sources
Principle: Use only sufficient, traceable context.
Authorized sources support the plan; material gaps are declared and block the affected step.
Go deeperHow to apply it
How to apply it
- Inventory sources and rights.
- State currency and limits.
- Exclude one unauthorized source.
Checklist
- Source tied to the case.
- Personal or authorized sources.
- Decision rule
- An inference never replaces a required source.
- Example
- Synthetic example: use selected notes and mark the missing week.
- Antipattern
- Granting broad access “just in case.”
- Expected evidence
- Source package, limits and gaps.
- Limit
- More context may increase exposure without improving relevance.
Limit tools and permissions
Principle: Authorize capabilities separately and by least privilege.
Reading, classifying, drafting and acting are distinct permissions with separate justification and revocation.
Go deeperHow to apply it
How to apply it
- Separate data, tools and actions.
- Mark prior approval.
- Remove one unnecessary permission.
Checklist
- Permission tied to step.
- Revocable permissions.
- Decision rule
- Anything undeclared remains prohibited.
- Example
- Synthetic example: read notes and draft; never send or delete.
- Antipattern
- Confusing available access with granted authority.
- Expected evidence
- Permission matrix and approval log.
- Limit
- Least privilege limits impact; it does not remove review.
Set restrictions and escalation
Principle: Every boundary needs an operational response.
Prohibitions, missing data and ambiguity must lead to stop, escalation or clarification, never improvised authority.
Go deeperHow to apply it
How to apply it
- List prohibitions.
- Assign escalation.
- Define how you will ask for help.
Checklist
- Observable stop.
- Stop remains under your control.
- Decision rule
- When authority is insufficient, stop before acting.
- Example
- Synthetic example: omit an unauthorized note and ask.
- Antipattern
- Writing “be careful” without testable conduct.
- Expected evidence
- Condition-response table and exceptions.
- Limit
- Escalation must not become automatic approval.
Approve the plan
Principle: Review the plan before any effect.
The gate corrects sequence, sources, permissions and reversibility while no environment has changed.
Go deeperHow to apply it
How to apply it
- Trace steps to sources.
- Trace steps to permissions.
- Issue your plan verdict.
Checklist
- Plan fits the brief.
- Plan verdict recorded.
- Decision rule
- An approved objective does not authorize every plan.
- Example
- Synthetic example: approve local drafting; reject publishing.
- Antipattern
- Reviewing only after the action occurred.
- Expected evidence
- Versioned plan and gate verdict.
- Limit
- An approved plan remains subject to exception and stop.
Execute with evidence
Principle: A useful run can be reconstructed.
Input, version, decisions, actions, gates, exceptions and result form the minimum reviewable record.
Go deeperHow to apply it
How to apply it
- Keep inputs and versions.
- Log decisions and gates.
- Keep your execution record.
Checklist
- Identifiable input and version.
- Execution reconstructable.
- Decision rule
- Fluency is not evidence of compliance.
- Example
- Synthetic example: notes, draft, changes and personal verdict.
- Antipattern
- Saving only the final output.
- Expected evidence
- Execution log and compliance table.
- Limit
- The log explains the case; it does not predict future performance.
Handle exceptions
Principle: Design the exception before it occurs.
A safe response states how to stop, who reviews and how to discard output or restore state.
Go deeperHow to apply it
How to apply it
- Classify the exception.
- Activate owner and gate.
- Choose your exception response.
Checklist
- Exception recorded.
- Recovery practicable.
- Decision rule
- Never invent permission to resolve an exception.
- Example
- Synthetic example: discard a draft that uses an excluded source.
- Antipattern
- Continuing to “finish” after traceability is lost.
- Expected evidence
- Exception, decision and recovery log.
- Limit
- Recovery does not remove the need to analyze cause.
Decide continuity
Principle: Continuity is a decision, not momentum.
Keep, adjust or retire depends on contract compliance, exceptions and changes required by another context.
Go deeperHow to apply it
How to apply it
- Review done and limits.
- Assign corrections.
- Choose keep, adjust or retire.
Checklist
- Verdict with reason.
- Continuity decided.
- Decision rule
- Do not widen scope or permission to rescue a failed pilot.
- Example
- Synthetic example: repeat the same case after fixing sources.
- Antipattern
- Turning one good case into general authorization.
- Expected evidence
- Continuity decision and versioned package.
- Limit
- A pass confirms that contract in that case, not every use.
Glossary
- Mission
- A bounded objective with context, restrictions and done.
- Gate
- A point where a person approves, corrects or stops.
- Least privilege
- Only the access or action required for the case.
- Exception
- A condition outside the expected flow that requires stop or escalation.
- Recovery
- A documented return to a safe state or disposal of the output.
Frequently asked questions
May an agent approve its own plan?
No. The plan is a proposal; approval belongs to the accountable person named in the contract.
What if a source is missing?
Declare the gap and stop the affected step. Never replace missing evidence with an assumption.
When should permissions expand?
Only for a concrete need, after risk review and new approval. A good pilot does not grant general authority.
What does a passed pilot mean?
It met done and its limits in that case; it does not validate the system for every context.
Expand autonomy only after the pilot passes; otherwise adjust or stop.
Design a bounded first agentic workflow: clear objective, minimum permissions, plan approval and evidence before expanding autonomy.
- Case, boundary, human owner and reviewer are written.
- Sources, permissions and prohibitions are clear.
- The plan was approved before the run.
- Exceptions and recovery were recorded.
- Continuity has a verdict, evidence and new authorization; this pilot grants no external effects.