subagent-dispatch

subagent-dispatch

Subagents are the vertical axis. Your main session is the horizontal axis. Getting the line between them right is what makes John work at hundreds of knowledge entries instead of melting your context.

Three tiers: inline subagent, one scale-out run, or batched runs

Before how to dispatch, decide the mechanism by the size and shape of the fan-out:

a few units, result needed in your context       → inline subagent (dispatch in waves yourself)
dozens–hundreds, uniform per-entry work           → one provider-native scale-out run
thousands                                         → batched scale-out runs per chunk-range,
                                                       all writing to the same event log

A scale-out run fans out the same subagents, keeps their results off the main context, adversarially cross-checks them, and returns only a summary. In Claude Code, use a dynamic workflow per [[vertical-workflows]]. In Codex, use native waves over the durable run ledger per [[codex-vertical-workflows]]. This skill covers the inline tier and the briefing discipline all engines share.

In Claude Code, check dynamic-workflow availability before the first fan-out and follow [[vertical-workflows]] when it is misconfigured or absent. In Codex, create the run ledger first and dispatch native waves per [[codex-vertical-workflows]]. Both branches emit the same events and reduce to the same checkpoint. Record the engine choice in PLAN.md either way. The rest of this skill is the inline mechanism and the briefing rules.

When to spawn a subagent

Three triggers, in order of clarity:

1. Per-entry work that fits one context window per entry but doesn't fit yours in aggregate. Classic case: 200 chunks to extract knowledge from. Each chunk is small; 200 of them through your context is not.
2. Work where you want a context firewall. Some tasks produce large intermediate state (a 50KB raw extraction) that you don't need in your context — you only need the digest. The subagent handles the raw; you see the summary.
3. Work that benefits from a tighter persona or cheaper model. A subagent can be given a narrow role ("you are a knowledge extractor; here is the schema; here is one chunk; emit entries to the event log and return a one-line digest") that focuses its output. John core delegates model selection to the active runtime; request the cheapest viable model when dispatching. Templates that need workerLLMs through external clients wire those themselves.