Two analysts used to burn two days building a single investment brief. A graph of Claude agents now does it in fifteen minutes, with citations on every claim.
The system is a LangGraph state machine, not a single mega-prompt. A planner decomposes the thesis, specialist workers retrieve in parallel, and a critic node gates the output, each with its own context, retries and exit conditions.
Treating orchestration as a typed graph means a failed source fetch or an unverifiable claim reroutes instead of poisoning the whole brief. The state is inspectable, so analysts can replay exactly how any sentence was produced.
Every sentence is bound to a source span at synthesis time, then independently re-checked by a critic agent. If a claim cannot be traced back to retrieved evidence, it is dropped or rewritten, not shipped.
That gate is what pushes citation accuracy to 98% and makes the briefs defensible. The desk reads research it can trust without re-verifying by hand.
| Component | Technology | Why |
|---|---|---|
| Agent orchestration | LangGraph | Typed, stateful graph with retries and conditional routing beats brittle prompt chains. |
| Reasoning core | Claude | Long-context synthesis and strict source-grounded writing with low hallucination. |
| Memory & recall | pgvector | Reuses prior research and dedupes context inside the existing Postgres stack. |
| Source layer | Python workers | Parallel async fetchers normalize 60+ heterogeneous feeds into clean chunks. |
| Research desk | Next.js | Streams briefs, citations and agent traces to analysts in real time. |
Async Python workers pull from 60+ filings, news and market feeds concurrently per run.
LangGraph coordinates planner, worker and critic roles with retries and conditional routing.
pgvector recalls prior research to enrich context and avoid duplicate work.
Each sentence is linked to the exact source span it was derived from.
A dedicated agent rejects any statement it cannot trace to evidence.
Next.js surfaces briefs, citations and agent traces to analysts as they generate.
“If an agent can't cite it, the desk never sees it.”
The Multi-Agent Research Desk turns a two-day, two-analyst grind into a fifteen-minute, fully-cited brief, and quadruples how much research the team ships. By treating orchestration as a typed LangGraph state machine and citation as a hard verification gate, it delivers speed without surrendering trust: 60+ sources per run, 98% citation accuracy, and an auditable trace behind every sentence. It is less a chatbot and more a research factory analysts can actually stand behind.