# Agent Teams Subagent Startup Convention

> Sub-skill of agent-teams: Subagent startup convention (+2).

- Skill: `vamseeachanta/agent-teams-subagent-startup-convention-2` (Agent Skill)
- Install (CLI): `npx skillmds@latest add vamseeachanta/agent-teams-subagent-startup-convention-2`
- Raw SKILL.md: https://api.skillmd.com/api/skills/vamseeachanta/agent-teams-subagent-startup-convention-2/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: vamseeachanta (https://skillmd.com/u/vamseeachanta)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/vamseeachanta/agent-teams-subagent-startup-convention-2

---


# Subagent startup convention (+2)

## Subagent startup convention


Set this as the **first Bash call** in every non-trivial subagent session:

```bash
export CLAUDE_SUBAGENT=1
```


## Hook behaviour by context


| Hook | Main session | Subagent | Reason |
|------|-------------|----------|--------|
| `session-review.sh` | Runs | Runs | Writes signals to `pending-reviews/` |
| `post-task-review.sh` | Runs | Runs | Captures findings to `pending-reviews/` |
| `consume-signals.sh` | Runs | Runs | Processes signals into session-briefing |
| `improve.sh` | Runs | **Skipped** | Narrow context produces low-quality output |
| `query-quota.sh` | Runs | **Skipped** | Quota tracking is main-session only |

The shared `.claude/state/pending-reviews/` directory is the passive message
bus. Subagents write signals via the lighter hooks; the main session's Stop
drains them all via `consume-signals.sh` → `improve.sh` at session end.


## Result


Main orchestrator stays lean and interactive. User can inject messages and
redirect subagents mid-flight. Learnings accumulate in the background without
adding any overhead to the orchestrator between Task completions.