Orchestration that survives contact, in brief
Lesson 1 set the rule: every seam between agents must justify itself. The AI Agent Teams track showed what seams can buy, a real team of agents arguing and checking each other. This lesson is the other side of the ledger: what seams cost, and the engineering that makes a coordinator-subagent system reliable when a step fails at two in the morning.
The capability: after this lesson, you can design a coordinator that decomposes a broad task without silent gaps, brief subagents that inherit none of the coordinator’s context, schedule independent work in parallel, and keep both failures and sources moving across agent boundaries in a form the coordinator can act on.
What the lesson covers. The coordinator decomposes, delegates, and aggregates; Anthropic’s orchestrator-workers pattern and its production research system ground the shape. The pivotal fact: a subagent starts with a fresh context window holding none of the coordinator’s conversation, and the only channel in from the coordinator is the delegation prompt, so the brief must carry an objective, an output format, source guidance, and boundaries. Decomposition gets tested in both directions, for gaps and for overlaps, because a too-narrow split produces a confident report about a smaller question. Independent subtasks spawn in parallel; dependent ones chain, with the earlier output pasted into the later brief. A failing subagent returns structure, what failed, what was attempted, what partial results survived, so the coordinator can retry, reroute, or proceed with the gap annotated. And findings travel with sources attached, because attribution dies in the middle hop otherwise.
Why this order. Lesson 4 ended by promising that tool contracts would return as agreements between agents. This lesson is that return: the delegation brief and its output format are the contract, now drawn between two agents instead of an agent and a tool.