// the take
DATAGEN is a LangGraph-orchestrated multi-agent pipeline: you hand it a CSV and a prompt, and it runs hypothesis generation, code-based analysis, visualization, and report writing agents in sequence to produce a research report. It's for people who want an automated, give-it-data-get-a-report workflow rather than a single LLM call that summarizes a dataframe.
Orchestration is an explicit LangGraph state machine (hypothesis -> human review -> processing -> quality review -> revision) rather than one opaque agent loop, so the flow can be reasoned about and interrupted at a defined point. Per-agent model configuration (agent_models.yaml plus individual AGENT.md/config.yaml files) lets you mix providers per agent, e.g. Claude for the code agent and Gemini for the note agent, and swap the whole config directory between dev and prod without touching code. CI enforcement goes past a badge: ruff, mypy, pytest, CodeQL and a dependency-audit workflow all run on PRs, and there are dedicated tests for agent failure bounds and loop bounds, which means someone specifically worried about runaway multi-agent loops rather than assuming it away.
The README itself warns that the agent system may modify the data being analyzed and tells you to back it up first, which is an unsettling admission for something pitched as a hands-off automated pipeline. A single report run fires off eight separate agents (hypothesis, process, visualization, code, search, report, quality review, note), each issuing its own LLM calls, so cost and latency compound through that chain and nothing in the docs gives a sense of typical token spend per run. Browser automation is wired in through a required ChromeDriver path, meaning Chrome/driver version pinning and headless setup friction, which gets worse in containers or CI. The maintainers' own 'Current Issues' section lists NoteTaker efficiency, overall runtime, and 'Refiner needs to be better' as open problems, so the revision loop at the center of the system isn't something even the authors consider solid yet.
We read the README, the directory tree and a key source file in November 2026. Nobody pays to be listed. How we judge →