Pick one primary pattern per organizational pain. Overlapping patterns need coordination — see multi-fleet.md.
flowchart TD
A[What hurts?] --> B{Don't know what agents exist?}
B -->|yes| C[Team Agent Registry]
B -->|no| D{Risky actions without approval?}
D -->|yes| E[Shared Inbox HITL]
D -->|no| F{Too many manager/worker handoffs failing?}
F -->|yes| G[Hierarchical Delegation]
F -->|no| H{One good agent, many teams want it?}
H -->|yes| I[Agent Clone & Fork]
H -->|no| J{Token bill surprise?}
J -->|yes| K[Fleet Budget Guard]
J -->|no| L{Compliance / incident review?}
L -->|yes| M[Cross-Agent Audit]
L -->|no| C
| Symptom | Pattern | Start with |
|---|---|---|
| "We have agents everywhere" | Team Agent Registry | F1 catalog only |
| Agents act without oversight | Shared Inbox HITL | F1 approve-only |
| Manager loop loses track of workers | Hierarchical Delegation | F1 typed handoffs |
| One agent should serve many teams | Agent Clone & Fork | F1 clone policy |
| Spend growing silently | Fleet Budget Guard | F1 caps only |
| "Who did this?" in incident | Cross-Agent Audit | F1 read-only audit |
- Team Agent Registry — inventory what exists
- Fleet Budget Guard — caps before scale
- Shared Inbox HITL — before any L2+ loops go unattended
| Combination | Rule |
|---|---|
| Registry + anything | Registry first; no new agents without manifest |
| Budget Guard + Hierarchical Delegation | Manager loop inherits team budget; workers inherit sub-caps |
| Clone & Fork + Registry | Clones must register as new IDs with forked_from |
| Audit + Inbox | Inbox approvals are audit events — link IDs |
If you do not yet have stable loops, start with loop-engineering first. Fleet patterns assume at least one loop pattern is understood.