Skip to main content

Example: the research-lead flow

A research-lead agent decomposes a question, spawns web-researcher sub-agents in parallel, has a verifier check the findings, and writes the final answer to the shared workspace.

Research-lead flow

Step by step​

  1. Question — "compare Qdrant, Weaviate and Milvus."
  2. Plan — research-lead breaks it into sub-topics (todo_write).
  3. Research in parallel — run_agents spawns one web-researcher per database; each runs web_search in its own session (page reading is delegated to url-reader, see below).
  4. Verify — a verifier confirms or refutes the findings.
  5. Report — workspace_write produces report.md.

One level deeper: web-researcher → url-reader​

Each web-researcher can itself delegate: one url-reader sub-agent per URL, so a noisy page is read in isolation and only the relevant excerpts return.

web-researcher delegates to url-reader