# Catenation > AI agents are great at tasks. Research isn't a task: it is hundreds of attempts at once, results built on results, and mistakes found a week later. Catenation runs it: people and AI agents on hundreds of campaigns at once, every result tracking what it rests on, agents attacking each other's work, and everything learned carried forward. Catenation gives frontier-model agents a persistent workspace for research that spans sessions: a fixed, versioned target; parallel, independent routes; exact scientific tools (Python and Wolfram, with independent replay); failed approaches kept on the record; typed evidence dependencies; and review by fresh reviewers before anything closes. The harness evolves: prompts, skills, memory and sub-agent roles are editable state, refined at turn boundaries with checksums, revision history and rollback. **The research record.** Catenation keeps the mathematical knowledge normally lost between discovery and publication. Alongside the final proof it keeps the intermediate derivations, failed constructions, counterexamples, changes in assumptions and verification evidence, linked so a researcher can see why a method works, where an alternative fails, and carry a technique to the next problem. Put agents to work on a paper and the campaign builds itself, every step recorded and formalized as it goes. The test is transfer: with the record, someone should solve a related problem, explain a necessary assumption or repair a broken argument better than with the paper alone. Key results: - **Nivat's conjecture, proved.** Posed in 1997 and open for 29 years, it was proved by a Catenation research campaign and independently verified by researchers in symbolic dynamics. Proof: [A proof of Nivat's conjecture](https://www.catenation.ai/research/nivat-manuscript.pdf) (Pranava Kumar, 14 September 2026, 21 pages), covering every finite alphabet and every k × n rectangle, with the equality case included. **The full proof is machine-checked in Lean 4**: [nivat-lean on GitHub](https://github.com/PranavaKCode/nivat-lean) proves `KumarNivat.nivat : FullNivat` in Lean 4.34.1 with Mathlib, using only Lean's standard axioms (propext, Classical.choice, Quot.sound). The library has 339 Lean source files, and all 923 axiom audits pass. - **Terminal-Bench Science 0.1: 72.8%** task resolution with GPT-6 Astra + Catenation, 8.2 percentage points above GPT-6 Astra's published 64.6%. The published comparison scores are Claude Fable 5.1 at 52.6% and Claude Opus 5 at 30.0%. - **ProteinGym: mean Spearman 0.53654** on the full benchmark (0.53204 held-out), above Aster's published 0.524. - The ensemble was selected on 125 development proteins and compared with the reproduced Aster formula on 61 held-out proteins: a +0.02519 gain, with a 95% protein-bootstrap interval of +0.01128 to +0.03770. - An independent evaluator replay passed across 217 assays and 2,465,767 variants. - **30 campaigns, 1,215 AI agents, at once.** These were 30 computational-biology research campaigns, with 37–44 agents on each, all on a local open-weight 27B model. - **Runtime built almost entirely in Rust.** - It uses 21× less memory than Claude Code at 1 active session, and about 22× less at 10. It uses 4.4× less than Codex CLI. - First frame comes in 53–110 ms, against 1.0–2.8 s for Claude Code. - **Pilots:** - Mantis Discovery, in the MIT CSAIL ecosystem: 20+ researchers and 15 manuscripts in 2026. - A Harvard Medical School pilot with Shaojun Pei on AI-guided tuberculosis vaccine development. Founder and CEO: Pranava Kumar. Contact: pranavak@mit.edu (pilots and research partnerships). ## Pages - [Homepage, as markdown](https://www.catenation.ai/index.md): product, benchmarks, the 30-campaign fleet, Nivat's conjecture, papers, team, direction and FAQ. - [Inside the research harness, as markdown](https://www.catenation.ai/research.md): architecture (programmable context, agents as calls, harness refinement, evidence dependencies, exact computation, research scouting) and the campaign gallery. - [Evidence and evaluation](https://www.catenation.ai/evaluation.html): sources, methods and records behind every number. - [Evaluation method](https://www.catenation.ai/evaluation-method.md): how the recorded runs and biology teams were measured. - [Technical notes](https://www.catenation.ai/technical-notes.html). ## Research - [A proof of Nivat's conjecture (PDF)](https://www.catenation.ai/research/nivat-manuscript.pdf) - [Lean 4 formalization of the proof (GitHub)](https://github.com/PranavaKCode/nivat-lean): the full conjecture, machine-checked, with build logs, axiom reports and a reproducible verification script. - [Nivat research record](https://www.catenation.ai/research.html#nivat-record): the proof's dependency graph from the campaign ledger, with 20 connected claims, 25 dependencies and 8 refuted shortcuts kept apart. - [Campaign gallery](https://www.catenation.ai/research.html#campaigns): the 30 computational-biology research questions, B01–B60. - [Mantis Discovery](https://mantis-discovery.vercel.app/): the pilot's researchers and manuscripts. ## Data (machine-readable) - [benchmarks.json](https://www.catenation.ai/benchmarks.json): Terminal-Bench Science, ProteinGym and runtime-efficiency results. - [evaluation-data.json](https://www.catenation.ai/evaluation-data.json): the recorded workload and system checks. - [nivat-history.json](https://www.catenation.ai/nivat-history.json): the Nivat campaign's source graph, counts and route history. - [bio-expanded.json](https://www.catenation.ai/bio-expanded.json): the 30-campaign biology fleet capture. - [proteingym-comparison.json](https://www.catenation.ai/research/proteingym-comparison.json) and [proteingym-replay.json](https://www.catenation.ai/research/proteingym-replay.json). ## Films - [Catenation — research that compounds](https://www.catenation.ai/media/links-film.mp4) (1:57), with [captions](https://www.catenation.ai/media/links-film.vtt). - [Catenation — product demo](https://www.catenation.ai/media/catenation-launch.mp4) (1:20), with [captions](https://www.catenation.ai/media/catenation-launch.vtt). - [The fleet: 30 campaigns at once](https://www.catenation.ai/media/fleet-1080.mp4) (0:40). ## Optional - [llms-full.txt](https://www.catenation.ai/llms-full.txt): this file plus both pages in full and every film transcript, in one document.