Summary: Orchestration that survives contact

A research crew can produce a confident, fluent, well-organized report that is missing every competitor with a physical storefront, without a single error firing anywhere. That is the signature failure of multi-agent systems: they fail in the seams, not in the sentences. This lesson is the engineering that makes coordinator-subagent teams reliable, the payoff of lesson 1’s rule that every seam must justify itself and of lesson 4’s promise that tool contracts would return as agreements between agents.

Core ideas

• The shape is Anthropic’s orchestrator-workers pattern: a central model “dynamically breaks down tasks, delegates them to worker LLMs, and synthesizes their results.” It earns its place only on tasks where the subtasks cannot be predicted in advance; predictable steps belong in code.
• Anthropic’s production research system is this shape at full size: a lead agent plans, spawns subagents that search in parallel, and synthesizes their distilled findings. Effort scales with the question, from one agent on a simple fact to several subagents on a comparison.
• The pivotal fact: a subagent inherits none of the coordinator’s work. It runs in a fresh conversation holding nothing of the parent’s; it gets its own system prompt, project memory, and tools, but no history. Intermediate tool calls stay inside it; only its final message returns. The only channel in from the coordinator is the delegation prompt itself.
• Isolation is a feature first: a subagent can read forty documents and none of them land on the coordinator’s desk. That is carry-versus-fetch, solved structurally.
• The price of isolation is the brief. Per Anthropic’s production guidance, each subagent needs an objective, an output format, guidance on tools and sources, and clear task boundaries. Boundaries should name the sibling subagent, exactly as lesson 4 named sibling tools.
• The most common orchestration bug is delegating as if the subagent were in the room: pronouns pointing at context that was never passed.
• Decomposition fails silently in two directions. Gaps: subtasks that add up to less than the task leave territory assigned to nobody, and nobody reports its absence. Overlaps: two briefs claiming the same ground buy the same facts twice and force reconciliation. Test both directions before spawning.
• Parallel when independent, serial when not. The test is one question per pair: does this brief need that one’s output? If yes, chain them and paste the earlier output into the later brief. If no, spawn together. Anthropic reports parallelism cut research time by up to 90 percent on complex queries.
• Parallelism buys back latency, never tokens. Multi-agent systems burn many times the tokens of a single chat; that is the entry fee for the seams.
• Failure is a finding. The two anti-patterns are the silent swallow (a subagent returns polished partial findings and never mentions what it could not do, laundering a failure into an invisible gap) and the kill switch (one failure discards everyone’s good work).
• The engineered seam carries structure: what failed, what was attempted, what partial results survived. With that, the coordinator can retry (transient), reroute (a different subagent or source), or proceed with the gap annotated in the final report. All three are legitimate; choosing is the coordinator’s job.
• Provenance dies in the middle hop. Every summarization strips sources by default, and multi-agent systems summarize at every seam. The fix is a rule in every output format: claim and source travel as one unit. Anthropic gives this a dedicated stage, a citation pass that pins every claim in the final report to its source.
• The blunt rule: a finding that crosses an agent boundary without its source arrives as a rumor.
• The full bill for a seam: tokens, gap risk, duplication risk, laundered failures, stripped provenance. All payable, with a briefing discipline, a partition test, a dependency test, a failure format, and a source rule.

What changes for you

The deep-research features in AI products you already use are coordinator-and-subagent systems. You now read their confident reports with the right three questions: what subtasks were actually assigned, what failed quietly, and can each claim be traced to a source? And when you build one, you know the discipline: brief for a fresh desk, test the split both ways, schedule by dependency, structure the failures, attach the sources. The next lesson takes the last step from surviving failure to knowing when not to proceed at all: escalation, human review, and the decisions an agent team must not make alone.