DeepResearch: a multi-agent research system

Give it a research question; nine cooperating agents plan sub-questions, search the web in parallel, cross-check every claim against at least two sources, and return a cited report with a confidence tag on each claim. It runs end to end on your own machine for $0.

Agent graph: the planner fans sub-questions out to parallel researcher, browser and summarizer branches; the fact checker and critic loop gaps back to the planner within a retry limit, then the code executor, writer and citation validator produce the cited report. — The LangGraph agent graph, drawn from the README.

Problem

A single research job can run for minutes and make dozens of LLM calls, so it has to run in parallel and survive a worker dying halfway through.
An agent that fails or finds nothing must not quietly ship an ungrounded report as if it were verified.

Approach

LangGraph’s Send() API fans sub-questions out in parallel, with per-node Postgres checkpointing.
Jobs run on Celery and RabbitMQ rather than FastAPI background tasks, which would tie up the API server and rule out horizontal scaling. With late acknowledgement, a crashed worker’s job is redelivered to another replica.
The critic can send gaps back to the planner, bounded by a retry-depth limit so it can’t research forever.
Ollama and DuckDuckGo are the defaults; every provider is swapped through .env and LiteLLM, so paid APIs are opt-in.
The code executor runs Python in a sibling container with no network and a 512 MB memory limit, so a crashed or hostile script can’t touch the worker.

What failed, and the tradeoffs

DuckDuckGo’s unauthenticated search rate-limits fast, so expect low-citation reports until a paid search key is set. The writer shows a “no sources retrieved” warning rather than hiding it.
Small local models occasionally write informally even with prompt constraints. That is a capability ceiling, not a code bug.
The worker’s Docker socket mount is a container-escape vector that this repo does not mitigate beyond network isolation. It is flagged rather than fixed, and a production deployment should put a scoped proxy in front.
The Kubernetes manifests are structurally complete, but only the Docker Compose path has been run against a live stack.
Crash recovery has no automated test.
The integration suite is flaky on Windows when pytest runs on the host, because cached async clients can outlive a test’s event loop.

Result

Every claim is tagged HIGH, MEDIUM, LOW or UNVERIFIED, and citations that can’t be reached are flagged rather than silently kept.
If an agent crashes mid-run, the job continues and the final report carries a visible warning.
A typical job costs $0 on the default local stack.

Stack

Python
LangGraph
Celery
RabbitMQ
FastAPI
Next.js
Postgres / pgvector
Docker