Design and Debug a Multi-Agent Research Pipeline

Why it matters

Section 1.2 makes the coordinator own delegation, failure handling and synthesis. You build that coordinator, then break it with a timeout and conflicting sources.

What to practise

Build a coordinator that delegates to at least two subagents, such as web search and document analysis; the exercise puts Task, an earlier name of the Agent SDK’s Agent tool, in allowedTools. Pass each subagent its findings in its prompt rather than relying on inherited context. Emit several delegation calls in one response and measure the latency against a sequential run. Give each finding a claim, excerpt, source and publication date, and check that synthesis keeps the attribution. Simulate a timeout: the coordinator should get the failure type, attempted query and partial results, and mark the coverage gaps. Then feed two credible sources with different statistics and check that both values survive, apart from the established findings.

How it fits

The coordinator is 1.2.1; it finishes with marked gaps, where 1.2.4 would re-delegate until the evidence suffices. In 1.3, allowedTools is 1.3.2, the explicit prompts 1.3.3, attributed findings 1.3.4, and the parallel calls 1.3.5. In 5.3, the timeout handoff is 5.3.2 and the annotated partial output 5.3.4. In 5.6, the finding structure is 5.6.1, the conflicting statistics 5.6.3, and the established-versus-contested split 5.6.4.