deep-research

deep-research

The "deep research" archetype Anthropic's multi-agent research system uses: one lead coordinator decomposes a question into subtopics, fans them out to researcher sub-agents called in parallel, then synthesizes their compressed, cited summaries into a report.

Piece	Lousho feature
Lead coordinator	createAgent({ instructions, subagents })
Parallel researcher fan-out	several task calls in one model turn (toolConcurrency defaults to 'unbounded')
Researcher sub-agents in clean contexts	subagents: { researcher } - each task sees only its prompt
Offline retrieval	a defineTool() search tool over a small in-memory corpus
Citations	[S#] markers the researchers return and the lead carries into the report and its Sources list

Run it

npx tsx examples/deep-research/index.ts                    # offline, scripted models
npx tsx examples/deep-research/index.ts "your question"    # offline, own question
OPENROUTER_API_KEY=... npx tsx examples/deep-research/index.ts  # live, openrouter/openai/gpt-4o-mini

Offline, both the lead and the researchers run on scripted mockModels: the researchers' search calls still execute for real against the corpus, so the run needs no network and no API key.

Test it

npx vitest run examples/deep-research

The tests check that one lead turn fans out >= 2 task calls, that the researcher runs actually overlap (stats.maxConcurrent), that the corpus was really searched, and that the report carries [S#] citation markers and a Sources list.