Build a research agent with sub-agents
⚠️
Preview — this feature is being built; the design below is the spec.
A lead agent breaks a research question into parts, hands each to a sub-agent with web search, and writes the final brief from their findings. Threads keep each sub-agent’s reading out of the lead’s context.
Design
| Agent | Model | Tools | Job |
|---|---|---|---|
research-lead | Amara Deep | Delegates only | Plan the questions, review findings, write the brief with sources. |
web-researcher | Amara S1.2 | syntic_web_search, syntic_web_fetch | Answer one narrow question with quotes and URLs. |
Set max_parallel_delegates to 3 or 4 so sub-questions run at the same time.
Lead instructions
- Split the question into 3 to 6 sub-questions that do not overlap.
- Give each delegate the sub-question and what form the answer should take: findings with a quote and URL for each.
- When results conflict, send a follow-up delegate to resolve the conflict rather than picking one.
- Write the brief with a source after each claim. Leave out anything no delegate could support.
Cost and time
Each delegate reads many pages; most tokens are spent there, on the cheaper model. Cap the session with a cost budget and the delegates with their own lower budgets. Watch per-thread usage with List threads to see where the money goes.
Quality checks
Grade briefs with a rubric that checks every claim has a source, and sample sources by hand to confirm the quotes exist on the page.
Was this page helpful?