Practice: Orchestration that survives contact
Self-check
Section titled “Self-check”Six short questions. Answer each in your head before opening the collapsible. Active retrieval is where the learning sticks.
1. What does a subagent inherit from its coordinator, and what is the only channel between them?
Show answer
None of the coordinator’s work. A subagent starts in a fresh conversation: it gets its own system prompt, project memory, and tools, but it never sees the coordinator’s history or earlier results. The only channel in from the coordinator is the text of the delegation prompt, and the only channel out is the subagent’s final message. Everything the subagent needs must be written into its brief.
2. Name the four elements of a delegation brief, per Anthropic’s production guidance.
Show answer
An objective, an output format, guidance on the tools and sources to use, and clear task boundaries. The boundaries should name the sibling subagents and draw the property line between them, the same way lesson 4 named a tool’s nearest sibling in its description.
3. Decomposition fails silently in two directions. What are they, and what is the test for each?
Show answer
Gaps and overlaps. A gap: the subtasks add up to less than the original task, so some territory is assigned to nobody, and nobody reports its absence. Test: what part of the original task belongs to no subtask on this list? An overlap: two briefs claim the same territory, duplicating work and producing conflicting versions of the same facts. Test: which two briefs could plausibly return the same fact?
4. When may two subagents be spawned in parallel, and what must happen when they cannot?
Show answer
Parallel only when neither brief needs the other’s output. If one does, they run serially, and the earlier subagent’s output gets pasted into the later subagent’s brief, because a fresh desk means it will never see that output any other way. Parallelism buys back latency but never tokens.
5. A subagent cannot complete its brief. What are the two anti-pattern responses, and what should it return instead?
Show answer
The silent swallow (return polished partial findings and never mention the failure, laundering it into an invisible gap) and the kill switch (abort the whole run, discarding every other subagent’s good work). Instead it returns structure: what failed, what was attempted, and what partial results survived, so the coordinator can decide what happens next.
6. What are the coordinator’s three legitimate moves when a subagent reports failure, and what kills attribution in a multi-agent system?
Show answer
Retry (the failure looks transient), reroute (point a fresh subagent at a different source), or proceed with the gap annotated in the final report. Attribution dies in the middle hop: every summarization strips sources by default, so unless the output format requires claim and source to travel as one unit, the final report asserts things nobody can check.
Try it yourself
Section titled “Try it yourself”Delegation exercise: brief for a fresh desk.
Pick one broad research or analysis task from your own work, something you would genuinely split across two or three helpers. On paper or in a note, work through four steps in order:
- Decompose it into two or three subtasks. Then run both tests: what part of the original task belongs to nobody, and which two subtasks could return the same fact? Fix the split before moving on.
- Write one full delegation brief for the trickiest subtask: objective, output format, sources, boundaries. Rule: no pronouns and no references to anything not stated inside the brief. Read it back pretending you know nothing else. Could you start work?
- Mark each pair of subtasks serial or parallel by asking whether one brief needs the other’s output. For every serial pair, note exactly what gets pasted forward.
- Write the failure report your trickiest subagent would return if its main source were unreachable: what failed, what was attempted, what partial results survived. Then decide, as coordinator: retry, reroute, or proceed with the gap annotated?
Keep the sketch. In the capstone you will orchestrate a real crew, and this exercise is the design half of that build.
Flashcards
Section titled “Flashcards”Q. What does a subagent inherit from its coordinator?
None of the coordinator’s conversation. It starts fresh, with its own system prompt, project memory, and tools, but no parent history. The delegation prompt is the only channel in from the coordinator; the final message is the only channel out.
Q. What are the four elements of a delegation brief?
An objective, an output format, guidance on tools and sources, and clear task boundaries. Anthropic’s production guidance, and the lesson’s contract between agents.
Q. Why is context isolation a feature before it is a burden?
A subagent can read forty documents and none land on the coordinator’s desk. Only the distilled finding returns. That is carry-versus-fetch (what an agent carries in context at all times versus what it fetches on demand) solved structurally.
Q. What is the most common orchestration bug?
Delegating as if the subagent were in the room: pronouns and references to context that was never passed. If it is not in the brief, it does not exist.
Q. What are the two silent failure modes of decomposition?
Gaps (subtask territory assigned to nobody, and nobody reports its absence) and overlaps (two briefs buying the same facts twice). Test the split in both directions before spawning.
Q. Why did Rosa's report miss every physical-storefront competitor without any error firing?
Too-narrow decomposition. The coordinator turned a broad market question into one narrow subtask, and the remainder was never assigned. A too-narrow split produces a confident report about a smaller question.
Q. When do two subagents run in parallel, and when serially?
Parallel when neither brief needs the other’s output. Serial when one does, and the earlier output must be pasted into the later brief.
Q. What does parallelism buy, and what does it never buy?
It buys back latency (Anthropic reports up to 90 percent less research time on complex queries). It never buys back tokens; multi-agent systems cost many times a single chat.
Q. What three things does a structured failure report contain?
What failed, what was attempted, and what partial results survived (with sources). It turns a failure into a finding the coordinator can act on.
Q. What are the coordinator's three moves after a subagent failure?
Retry (transient failures), reroute (a different subagent or source), or proceed with the gap annotated in the final report. Swallowing the error and killing the run are the anti-patterns.
Q. What happens to attribution in the middle hop?
It dies. Summarization strips sources by default, and every seam summarizes. Claim and source must travel as one unit through every hop.
Q. What is a finding without its source, one boundary later?
A rumor. One hop after that, unchecked background truth. Require sources in every output format so the final synthesis can cite what it asserts.