The architecture: orchestrator-worker

The system uses the orchestrator-worker pattern, built from three kinds of component.

The Lead Researcher agent is the orchestrator. When a query arrives, it analyzes the request, decides on an overall strategy, and writes its plan into memory. That last detail matters: large research tasks can exceed the model's context window, and the system persists context when conversations pass 200,000 tokens so the plan and findings are not lost when tokens run out. This is memory used as a safeguard against context rot.

The subagents are specialized workers spawned by the lead. Each gets a specific assignment (explore a particular company, a time period, a technical detail), and each runs in its own context window with its own tools and its own exploration path. Because they run in parallel and do not interfere with each other, they cover far more ground than a single agent could. They act as intelligent filters: they gather a lot, then condense their findings before sending them back to the lead.

The Citation Agent is a dedicated final component. After the research is gathered, it checks every claim against its sources, matches citations correctly, and makes sure the report is traceable. This guards against hallucination and against attributing a fact to the wrong source.

The full flow: the user submits a query, the lead plans, subagents are spawned to search in parallel, the lead gathers and synthesizes their results and decides whether more work is needed (spawning more subagents or refining the strategy), and once enough is collected the Citation Agent verifies sourcing before the final report is returned. The system also uses parallelism at two levels: the lead spawns 3 to 5 subagents at once, and each subagent runs multiple tool calls in parallel, cutting research time by up to 90 percent for complex queries.