FARS: Fully Automated Research System

TL;DR

FARS is Analemma’s “Fully Automated Research System” — a multi-agent pipeline (Ideation → Planning → Experiment → Writing) that runs continuously and autonomously on a 160-GPU cluster to produce short AI research papers, one every ~2 hours, at ~$1K/paper. The first public deployment was a 228-hour livestream that produced 100 papers from 244 hypotheses, burning ~11.4B tokens. The framing is “AI-for-AI” research: optimize for hypothesis throughput rather than academic-publication conformance, explicitly admit negative results as valid output, and use scale + public observability to outsource quality evaluation to the broader research community.

Key claims

FARS is structured as four specialized agents — Ideation, Planning, Experiment, Writing — coordinating through a shared file system that doubles as workspace and persistent memory; no direct agent-to-agent messaging is required [§Method].
The Ideation agent takes a multi-direction prompt document, performs literature review against open-access papers + public GitLab repos, and emits hypotheses; only hypotheses that pass automated review are forwarded downstream [§Method].
The Experiment agent has access to a 160-GPU NVIDIA cluster encapsulated as a tool, plus a range of model-inference endpoints for data synthesis, agent design, and LLM-as-Judge subtasks [§Method].
The first public deployment ran for 228 hours 28 minutes, produced 100 short papers from 244 hypotheses, and consumed ~11.4B tokens [36kr coverage].
The system targets minimal, composable contributions rather than full academic papers — each output is a single hypothesis + its validation, may explicitly report negative results, and is not required to conform to standard length/structure conventions [§Building an AI Research System from First Principles].
The motivating design constraint is that meaningful evaluation of an automated research system requires scale, because the output distribution varies widely across topics/methods/quality and no single team can reliably assess it in isolation [§A Live and Large-Scale Deployment].
Recommended starting research direction at deployment time is RLVR (Reinforcement Learning from Verifiable Rewards), though the system is free to explore other topics [§Method].
Pre-arXiv release pipeline still requires manual review by at least three internal researchers with 5+ years experience; arXiv policy does not permit AI authorship so distribution is being explored through other indexable channels [§Building an AI Research System from First Principles, post-deployment update].

Method

FARS instantiates the research pipeline as four role-specialized agents executed in sequence, with the shared filesystem as the only coordination substrate. Ideation reads a topic-direction document and produces hypotheses by combining literature retrieval (open-access papers, public GitLab repos) with LLM-based generation. Each hypothesis passes through an automated review gate; survivors get a structured project directory. Planning fills in an experiment design; Experiment runs it against the encapsulated 160-GPU cluster + inference endpoints; Writing produces a short paper from the resulting artifacts. Because all stages read/write the same project directory, FARS can run many concurrent projects through the pipeline in parallel — the system observed in deployment shows a project queue with stages in flight simultaneously.

The deployment philosophy is unusual: instead of internal-evaluation-then-release, Analemma livestreams the entire research session publicly, publishes all papers + code to a dedicated GitLab account, and invites external review. Pre-arXiv submission is gated by a 3-reviewer human pass; everything else is observable in real time. After the first deployment, Analemma is piloting Stanford’s Agentic Reviewer (paperreview.ai) for automated review of FARS outputs, with results published but flagged as non-official.

Results

228 h 28 min 33 s continuous public operation; 244 hypotheses → 100 completed short papers; ~11.4B tokens; ~$1K per paper; ~2 h per paper.
The first deployment concluded March 3, 2026; all papers + code remain hosted at analemma.ai/fars.
No quantitative quality assessment is yet published — Analemma’s stated next step is a systematic human review with external reviewers, plus exploration of indexable distribution channels (since arXiv policy bars AI authorship).
The blog explicitly does not report SOTA on any benchmark; the deliverable is the system, and the empirical quality of the 100-paper corpus is deferred to the planned external review.

Why it’s interesting

This is the wiki’s first datapoint on end-to-end research automation — the agent-system substrate (multi-agent + shared workspace + LLM-as-Judge) is the same family as Tool-Use Agents and Agentic Software Engineering, but the optimization target is “produce a research contribution” rather than “fix a bug” or “answer a query.” Two observations Luma should track: (a) the closed-loop synthesis pattern that dominates agentic-SWE (model → synthesizer → next model) hasn’t yet been reported here, but the architecture is structurally compatible — the 100-paper corpus is exactly the kind of dataset one would re-inject into the next training round; (b) the throughput economics (~$1K, ~2 h per paper) imply that the bottleneck for AI-for-AI research is shifting from compute toward evaluation — there is now more research being produced than reviewers can read, which is also Analemma’s explicit motivation for livestreaming.

Dejia’s framing in the originating Slack post is the operative one for Luma: “most papers look incremental, but there will soon be more need for infographics generation.” If automated research systems produce papers at industrial throughput, the downstream demand is for figure-generation pipelines that can keep up — exactly the workflow PaperBanana: Automating Academic Illustration for AI Scientists targets. FARS is the upstream half (paper generation); PaperBanana is the downstream half (figure generation). Worth watching whether Analemma’s pipeline integrates a PaperBanana-style figure agent for the Writing stage.