↓ Skip to main content
tomrochette.com
  1. Agents/

Automated research

Author
Tom Rochette
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: in progress / Readability: 3 / 6 minutes read (~1079 words)
agents 48 automated-research 32
Table of Contents

Where the research loop itself runs autonomously: the labs’ science programs, the productized research agent, and the community-built loops from markdown skills and idea-to-paper pipelines to multi-agent evolution, metric-judged tree search, and financial data, plus the formal-proof engines pointed at the hardest open problems.

  • Agent Laboratory - Samuel Schmidgall’s human-piloted idea-to-report framework whose own evaluation scored its papers 4 of 10 against the accepted-NeurIPS average, with AgentRxiv as the shared preprint server, dormant since 2025-08-20, 5,900 stars as of 2026-10-11.
  • Agon - the MIT Claude Code plugin running adversarial producer-critic factories from a one-line topic to experiments and a paper draft, with a candid failure-mode taxonomy.
  • AI-Researcher - HKU Data Intelligence Lab’s NeurIPS 2025 idea-to-paper system with its own Scientist-Bench, the sub-family’s largest community and no license file, dormant since 2025-10-16, 5,802 stars as of 2026-10-11.
  • AIDE ML - Weco AI’s MIT reference build of the AIDE tree-search agent that drafts, debugs, and benchmarks ML code toward a user metric, the mechanism’s published ancestor, 1,572 stars as of 2026-10-13.
  • AiScientist - AweAI’s File-as-Bus research lab whose paper track reproduces published work on PaperBench and whose MLE track pushes a metric across 24-hour runs, dormant since 2026-07-21, 146 stars as of 2026-10-11.
  • AlphaProof - DeepMind’s Lean reinforcement-learning solver, from IMO silver in 2024 to officially graded IMO gold via Deep Think in 2025.
  • Anthropic Claude mathematical research - the Claude Code subagent loop that raised the zeta zero bound to 67.2 percent and formalized Fermat’s Last Theorem in 11 days.
  • Arbor - the RUC NLPIR hypothesis-tree framework where a long-lived coordinator dispatches short-lived executors across git worktrees and only merges gains that clear a margin on held-out data, 1,106 stars as of 2026-10-12.
  • ARIS - wanshuiyin’s MIT markdown-skill research loop where the executor drives and a reviewer from a different model family judges every artifact across eight harnesses, 17,367 stars as of 2026-10-13.
  • AutoContext - Grey Haven’s Apache-2.0 recursive self-improvement harness whose promotion gate keeps only context and training changes that beat the incumbent on matched, replayed trials, 1,308 stars as of 2026-10-14.
  • AutoResearchClaw - the aiming-lab 23-stage idea-to-paper pipeline with multi-agent debate, four-layer citation verification, and seven human-in-the-loop modes, dormant since 2026-08-19, 14,618 stars as of 2026-10-11.
  • AutoSci - the memory-centric full-lifecycle research platform with SciMem, SciFlow, SciDAG, and SciEvolve, stable on Claude Code with Codex and OpenCode previews, 1,698 stars as of 2026-10-11.
  • CORAL - the Human-Agent Society multi-agent evolution framework where a manager partitions coding agents into grader-isolated worktree islands and shares wins through persistent memory, COLM 2026, 1,063 stars as of 2026-10-11.
  • DeepAnalyze - RUC DataLab and Tsinghua’s trained 8B agentic model that runs the data-science loop from raw files to analyst-grade reports with no external judge, 4,678 stars as of 2026-10-11.
  • Dexter - Virattt’s open-source autonomous financial research agent that plans, executes, and self-validates tool runs over live financial statements with a JSONL audit trail, 27,652 stars as of 2026-10-11.
  • EvoMap AutoResearch - the two-stage workflow whose deliverable stops at an evidence package of runs, critic reports, and blind reviews, its launch shadowed by a forged-commit incident that briefly showed Karpathy as a contributor, 3,016 stars as of 2026-10-12.
  • FutureHouse Robin - the nonprofit’s Nature-published wet-lab loop that proposed and validated ripasudil for dry AMD, with people at the bench and its own ablations as the critique, 733 stars as of 2026-10-11.
  • GPT Researcher - the Apache-2.0 open deep-research agent running the plan-search-filter-write loop on any LLM since 2023, cited in 150-plus papers as the baseline, 30,060 stars as of 2026-10-13.
  • Harmonic Aristotle - the free agentic theorem prover with a Lean-checked Erdős result, now aimed at software correctness, judged and contested on FormalQualBench.
  • Kapso - Leeroo’s MIT campaign framework where long-running agents optimize AI and data systems toward a declared evaluator, banking evidence-priced lessons between campaigns, with results on MLE-Bench, ALE-Bench, RelBench, and the IOAI 2026 AI Model Track, 123 stars as of 2026-10-13.
  • Karpathy Autoresearch - Karpathy’s three-file keep-or-revert loop where an agent edits a nanochat training script overnight and the validation curve judges every experiment, the family’s founding artifact, 97,798 stars as of 2026-10-13.
  • Local Deep Research - the community’s MIT local-first research agent running an agentic search loop over arXiv, PubMed, and your own encrypted document index on local models, 9,187 stars as of 2026-10-13.
  • Math Inc. Gauss - the autoformalization agent behind Strong PNT and the sphere-packing proof, with the comparator-audited OpenGauss harness, as of 2026-09-13.
  • Mole - Lajos Deme’s Go deep-research agent with a reserve-settle budget ledger that stops runs at the ceiling you set and verbatim-quote checks that drop claims the source never said, 314 stars as of 2026-10-12.
  • NanoResearch - OpenRaiser’s nine-stage idea-to-paper engine that submits experiments to local or SLURM GPU clusters and writes the paper from the logs those runs return, 1,340 stars as of 2026-10-11.
  • OpenAI Deep Research - the productized web-research agent, the breadth-first loop with no machine judge behind it.
  • OpenAI for Science - the lab program from FrontierMath’s first open-problem solve to the disputed Navier-Stokes claim, as of 2026-09-13.
  • OpenResearch - alphaXiv’s MIT local-first workspace that turns Claude Code, Codex, OpenCode, Cursor, or Antigravity into research agents with an experiment tree and an autonomous loop, 6,937 stars as of 2026-10-13.
  • Pion - Andon Labs’ closed research preview where persistent agents run a real business with payment tools, the Vending-Bench lineage made product.
  • Sakana AI Scientist - the idea-to-paper ancestor whose v2 produced the first fully AI-generated paper to pass a workshop peer review, withdrawn from publication by design, dormant since 2025-12-19, 7,294 stars as of 2026-10-11.

Its members are compared on shared rows in the Automated Research Feature Matrix.

Changes
#

  • 2026-09-13 - Added AlphaProof.
  • 2026-09-13 - Added Anthropic Claude mathematical research.
  • 2026-09-13 - Added Harmonic Aristotle.
  • 2026-09-13 - Added Math Inc. Gauss.
  • 2026-09-13 - Added OpenAI Deep Research.
  • 2026-09-13 - Added OpenAI for Science.
  • 2026-09-16 - Added Pion.
  • 2026-09-27 - Added Agon.
  • 2026-10-05 - Added OpenResearch.
  • 2026-10-06 - Added FutureHouse Robin.
  • 2026-10-07 - Added Local Deep Research.
  • 2026-10-07 - Added GPT Researcher.
  • 2026-10-10 - Added DeepAnalyze.
  • 2026-10-10 - Added Karpathy Autoresearch.
  • 2026-10-10 - Added ARIS.
  • 2026-10-10 - Added AutoResearchClaw.
  • 2026-10-10 - Added Dexter.
  • 2026-10-11 - Added Kapso.
  • 2026-10-11 - Added Sakana AI Scientist.
  • 2026-10-11 - Added Agent Laboratory.
  • 2026-10-11 - Added AIDE ML.
  • 2026-10-11 - Added AI-Researcher.
  • 2026-10-11 - Added AiScientist.
  • 2026-10-11 - Added AutoSci.
  • 2026-10-11 - Added CORAL.
  • 2026-10-11 - Added NanoResearch.
  • 2026-10-12 - Added Arbor.
  • 2026-10-12 - Added EvoMap AutoResearch.
  • 2026-10-12 - Added Mole.
  • 2026-10-14 - Added AutoContext.