Design and Debug a Multi-Agent Research Pipeline

Why it matters

A risk register for an agent system is guesswork until you have watched one failure travel through it. This exercise has you cause failures on purpose and trace where they land.

What to practise

Build a coordinator that delegates to at least two subagents, passing each its findings in the prompt rather than relying on inherited context, and fan them out in parallel with several subagent calls in one response. Give each finding a claim, an evidence excerpt, a source and a date. Then simulate a subagent timeout: check that the coordinator receives a structured error with the failure type, the attempted query and any partial results, and that the report names its coverage gaps. Feed in two credible sources that disagree and check that both values survive with attribution. The exercise names the subagent tool Task in allowedTools; the Claude Agent SDK has since renamed it Agent.

How it fits

The timeout is a dependency failure of the kind 5.2.2 traces across a workflow. Whether it is caught or becomes a confident, incomplete answer is the detectability that 5.2.4 ranks. Proceeding on partial results with an annotated gap is a fallback, and 5.2.5 asks who owns that decision and accepts the residual risk.