subagent-delegationlisted
Install: claude install-skill alectivism/organization-ai-skills
# Subagent Delegation
The session model is the expensive orchestrator: it plans, judges, and writes the final output. Gathering, reading, checking, and bulk work belong on cheaper worker subagents. On Anthropic's reported orchestrator-worker benchmark this split kept 96% of the top model's accuracy at 46% of the cost on BrowseComp; OpenAI's GPT-5.6 family is built for the same pattern.
## The roles, and the model to pin
Pick the column for the platform you are running on. Pin model and effort per worker; that choice is the cost policy.
| Role | Claude | GPT-5.6 | Use for |
|---|---|---|---|
| `researcher` | Sonnet / medium | Luna / high | web and file research; 3+ searches or sources |
| `summarizer` | Sonnet / low | Luna / medium | long docs, threads, transcripts into a tight brief |
| `bulk-worker` | Haiku / low | Luna / low | repetitive, fully specified work across many items |
| `verifier` | Opus / high | Sol / high | the high-stakes gate: verify drafts, numbers, and claims before sending or acting |
| `planner` | Opus / high | Sol / high | hard planning, architecture, or reasoning (see the escalation rule below) |
Spawn a worker by setting its model explicitly; never let an ad-hoc subagent inherit the orchestrator's expensive model. If none of the roles fits, use a general-purpose subagent with the model set to the cheap tier (Sonnet, or Luna).
## Delegate down by default; escalate up only when it helps
- **Down (researcher, summarizer, bulk-worker)** is the defau