Agent Delegation Framework
Not all delegations are equal. A 7-level spectrum helps you match the right amount of autonomy to the right situation — preventing both micromanagement and unsupervised failures.
Key Takeaway
The right delegation level depends on two axes: agent capability (has it done this before?) and task risk (is the outcome reversible?). High capability + low risk = high autonomy. New task + high stakes = tight oversight.
The 7-Level Delegation Spectrum
| Level | Name | Description | Agent Instructions |
|---|---|---|---|
| 1 | Tell | Execute exactly as specified | "Run this exact command. Report the output." |
| 2 | Sell | Execute with context on why | "Run this command. Here's why it matters: [context]" |
| 3 | Consult | Gather input, orchestrator decides | "Research options A, B, C. Present findings. I'll choose." |
| 4 | Agree | Discuss and decide together | "Analyze the problem. Propose solutions. We'll decide together." |
| 5 | Advise | Agent decides, orchestrator can veto | "Choose the best approach. Tell me before executing." |
| 6 | Inquire | Agent decides, reports after | "Handle this. Let me know what you did." |
| 7 | Delegate | Full autonomy, no report needed | "This is your domain. Handle it completely." |
The Decision Matrix
Choose the delegation level by mapping agent capability against task risk:
TASK RISK
Low High
┌─────────┬─────────┐
High │ Level │ Level │
AGENT │ 6-7 │ 4-5 │
CAPABILITY│ (Trust) │(Collab) │
├─────────┼─────────┤
Low │ Level │ Level │
│ 3-4 │ 1-2 │
│(Develop)│(Direct) │
└─────────┴─────────┘Key Questions
- Has this agent successfully done this before? → Higher level
- Is the outcome reversible? → Higher level acceptable
- Is there time pressure? → May need to match to capability
- Is this a learning opportunity? → Consider level 3-4
Practical Examples
Level 1-2 (Direct): First-time sensitive task
"Run git push --force origin main. This will overwrite remote history. Only do this because we need to remove the accidentally committed API key. Report the output."
Level 3-4 (Develop/Collaborate): Research with decision
"Research the top 3 options for serverless databases (Supabase, PlanetScale, Neon). Compare on: pricing for our use case, PostGIS support, edge compatibility. Present your analysis and recommendation. I'll make the final call."
Level 5-6 (Trust): Routine task, proven capability
"Write the daily memory file for today. You know the format and you've done this 30 times. Handle it and let me know when done."
Level 7 (Full Delegate): Agent's domain
"Monitor heartbeats and handle routine checks. This is your territory."
FAQ
How do I increase an agent's delegation level over time?
Start at level 3-4 for new task types. If the agent succeeds 3+ times consistently, move up one level. If it fails, drop back one level and add more context to the task spec. Think of it like training a new team member — start supervised, earn autonomy through demonstrated competence.
What if the agent is at level 7 but makes a mistake?
Drop to level 5 (advise — tell me before executing) for that task type. Analyze the failure. Update the spec or constraints. Move back to level 7 only after the root cause is addressed and the agent demonstrates the fix works. Never keep an agent at high autonomy after a failure in that domain.