Agent Teams
Sub-Skills
- Core Constraint
- Workspace Work Profile (Updated 2026-03-10)
- Activation
- Decision Matrix: Team vs Sequential
- 1. Create team (+5)
- DM (default — always prefer) (+2)
- Agent Types for Common Tasks
- Work Queue Integration
- Subagent startup convention (+2)
- Idle State
- Related
Iron Law
No task shall spawn more than 3 concurrent subagents (MAX_TEAMMATES=3) — no matter how parallelizable the work appears.
Rationalization Defense
| Excuse | Reality |
|---|---|
| "This task has 5 independent parts — 5 agents would be faster" | More agents means more coordination overhead, more context conflicts, and more merge failures. The limit is empirically derived. |
| "I'll just use 4 this one time" | The limit exists because every team that exceeded 3 hit coordination failures. There is no safe "just one more." |
| "These agents don't need to coordinate" | Even "independent" agents contend for git locks, file writes, and context. Zero-coordination is a myth in a shared repo. |
Red Flags
These phrases signal you are about to violate the Iron Law:
- "we could parallelize this into N agents" (where N > 3)
- "one more agent won't hurt"
- "these are fully independent — no coordination needed"
- "the limit is conservative for this case"