Sub-Agent Architecture

A sub-agent is a scoped worker spawned by a main agent; it does not share a context window with siblings. Communication is mediated by the main agent — "you essentially communicate with the sub agents only via your main agent. And you report back once" (faye-zhang).

Contrast with agent swarms

Swarms let agents talk peer-to-peer. Zhang's team explicitly rejected this for post-training because: - Context windows "quickly expand exponentially" as every agent re-receives every sibling's output. - Orchestration rigidity produces the hot-celebrity-orchestration-problem.

Why this matters

The "communicate only via the main agent" constraint is exactly the harness discipline Lopopolo (harness-engineering) and Chase/Partee (harrison-sam-general-purpose-agents-2026) advocate for scoping. Sub-agents are the microservice form of agentic computing — they only work if your orchestrator treats them as stateless and the main agent owns the global context.