Multiple agents and delegation

5.1 Who may delegate to whom

Declare it in the agent’s frontmatter:

---
model: deepseek/deepseek-flash
subagents: [task]
---

This means “this agent may hand work to task”.

★ Referencing an unknown agent, or declaring a cycle, is a startup error — never a silent truncation.

5.2 The depth cap

The cap is decided by agent.max_depth in server/config.yaml together with the same key in policy/limits.yaml, and the stricter one wins. The default 1 means “only the entry agent may delegate; a sub-agent may not spawn children”.

For a deeper delegation chain:

# policy/limits.yaml
agent:
  max_depth: 3

★ A declared topology that exceeds the cap is a startup error, not “truncate to the cap and carry on”. The latter would make the depth protection meaningless.

5.3 Several agents in one session

Independent work can be dispatched in parallel. Usage is accounted per agent (usage.byAgent in /v2/sessions/{id}/messages), so it is obvious who spent what.

In the interface:

• The agent badges in the session list filter the history by agent.
• A sub-agent’s output in a message carries an attribution marker, so you can see who produced it.