Antinuous

100x more experiments, 1,000x cheaper.

Autonomously run research, experiments, and simulations thousands at a time, so hours of sequential runs take minutes. Antinuous builds the compute, orchestration, simulation, and artifacts underneath, for our agent or yours.

Sign up
Backed byY Combinator
An agent swarm runs experiments on a compute cluster and keeps what it finds. Groups of agents draw together and break up again, a hotspot of machines pulses where each group sends its work, and the findings turn into the artifacts they hold: an aircraft in airflow, a crystal lattice, a price chart, a truss bridge under load, a cited document, a fusion plasma, a circuit board, a strand of DNA, a plate under stress, and a dose-response chart.
Used by researchers, analysts, and engineers at these organizations
NVIDIA
Stanford
Morgan Stanley
Columbia University
Vest
OpenAI
Georgia Tech
Uber
Wharton
Datadog
Brown University
SpaceXSI
University of Pennsylvania
LinkedIn
Macquarie

Four layers

Antinuous is built in four layers. Our agent runs on them, and your agent can use them too.

  • A cluster of computers, most of them running work, with more being added at the edge as demand grows
    Compute

    Elastic compute for research. Each run gets its own computer, and large jobs split across thousands of workers, so hours of sequential runs take minutes.

  • A lead agent starts a swarm. Each round, weaker branches are dropped after review, and the rest return their results.
    Orchestration

    Agents work together as swarms. A large question is split across many agents that finish 3.5x sooner than one, working in rounds and checking sources.

  • Four simulations side by side: a loaded plate with a hole coloured by stress, a crystal of atoms, simulated price paths that end up or down, and a storm on a weather grid
    Simulation

    Agents test ideas in simulation instead of guessing, so only the best of thousands reach a prototype, across engineering, science, medicine, finance, and earth science. They use our own kernels and solvers alongside established ones.

  • A notebook, a chart, a deck, a molecule, a table, a CAD bracket, and a sequence, each in its own form
    Artifacts

    Results arrive in the formats your field already uses, such as notebooks, documents, decks, charts, maps, tables, CAD models, and protein structures. Each value links to its source or code.

How our agent uses it

A question goes to an agent. When massively parallel computation can speed up the work, the agent starts a swarm that runs experiments, reads sources, and returns artifacts 3.5x faster than one agent, and up to 19x. Rounds repeat until the results hold up.

  1. A question

    • Plain words

      Describe an idea in plain words, or send it from a coding agent.

    • Smaller questions

      Antinuous breaks it into smaller questions, with one search or simulation for each thing to study.

    • On a schedule

      Scheduled reruns run as a swarm, 3.5x faster.

    A question branches into smaller questions and tasks
  2. Agent swarms

    • One part each

      When a request can be split and run in parallel, the agent starts a swarm, and each agent takes part of the question.

    • In parallel

      They search and plan at once, 3.5x faster than one agent.

    • Reports back

      Each agent reports back what it found.

    A lead agent splits the question across teams of researchers that hand work to each other
  3. Compute clusters

    • Its own computer

      Every experiment, simulation, and analysis runs as code on its own computer, so 3 hours of sequential runs take 8 minutes.

    • Reused work

      Antinuous plans each batch and reuses work across runs, so a rerun skips work already done.

    • Reruns

      You can read and rerun any of them.

    Agents send work to many computers, which return results
  4. Sources

    • Web and databases

      Agents search the web, science databases, and public records.

    • Full papers

      They read full papers and flag retracted ones.

    • Your files

      They open your PDFs, spreadsheets, CAD models, and data files.

    • A link for each fact

      Each fact links to its source, from up to 250 per run.

    Papers, web pages, tables, files, and records collected together, each in its own form
  5. Artifacts

    • The right format

      Results come back as a sheet, notebook, chart, deck, or draft, whichever fits the question. Each value links to its source or code.

    • Exports

      Decks export to PowerPoint, PDF, or Google Slides.

    • Your approval

      Drafts and other actions wait until you approve them.

    A sheet, notebook, chart, deck, and draft laid out together at different sizes
  6. Rounds

    • Plan, gather, run, review

      Each round plans, gathers evidence, runs simulations, and reviews the results.

    • Weak candidates dropped

      Weak candidates are dropped each round, so each round runs fewer jobs than the last.

    Each round drops weak candidates and keeps the rest

What agent swarms can run

Each run gets its own computer with simulation and engineering tools installed. Agents run code there instead of guessing, and split large jobs across many workers, so hours of sequential runs finish in minutes.

  • One agent splits work across a pool of workers that return results
    Worker pools

    Independent variants, seeds, and configurations split across many workers at once. GPU jobs run on clusters directly through Antinuous, or on your own Modal account.

  • A formula car in a frontal crash into a barrier, coloured by stress from grey to red in the nose and front wing
    Structures and crash

    Static, frequency, buckling, and heat transfer analysis, plus explicit crash and impact simulation.

  • Air flowing past an aircraft, coloured by flow speed from slow to fast
    Fluids and aerodynamics

    Parallel CFD, wing and airfoil screening, and combustion and fluid property models.

  • A circuit board with chips, connectors, capacitors, a battery pack showing charge and temperature per cell, and a signal trace
    Systems, circuits, and batteries

    Multibody, hydraulic, and thermal system models, circuit and board simulation, and battery cell models.

  • A flange drawn as a part on one side and meshed on the other, with finer elements near holes and edges
    CAD and meshing

    Parametric parts and assemblies with STEP files and drawings, meshed for analysis. Models open in 3D.

  • A molecule from a chemistry database beside the papers that describe it
    Science databases

    Records from UniProt, PDB, AlphaFold, ChEMBL, PubChem, ClinicalTrials.gov and NCBI Gene, and full papers checked for retraction.

Computed values are saved with the method and inputs that produced them, and researched values link to their source.

Research primitives for any agent

Claude, ChatGPT, Claude Code, Codex, Cursor, and other MCP clients call the same tools the app uses. A local or cloud agent can start our agent, steer it while it runs, and read what it finds. It can also start a swarm directly and change results itself.

On its own, an agent runs experiments one after another. The same agent calling Antinuous runs thousands at once, and finishes 13x faster.
Start and steer our agent
  • Start research

    Starts a run from a question and returns a link to the run right away while the work runs.

  • Steer a run

    Sends follow-ups that queue while it runs, checks its status, or stops it and keeps the results saved so far.

Read what it finds
  • Read results

    Reads tables, rows, conversations, and documents, with how many sources support each cell.

  • Read a value's evidence

    Returns a value with its sources, the reasoning behind it, and a confidence.

Run work directly
  • Start a swarm

    Starts a swarm of agents in a thread, each on its own computer, and steers or stops each one.

  • Change results

    Changes results directly, without a follow-up.

Agents that speak A2A can start and continue research the same way. Every call follows the same access rules as the app.

See the platform

Where it's used

What people run on the platform. Pick a field to see examples.

A DNA double helix beside a cross-section of a fusion tokamak, its plasma hottest at the core

Applied sciences

Run thousands of experiments in silico at once, so only the best candidates reach the bench.
A truss bridge carrying a truck, coloured by stress from grey to red in the most loaded members

Engineering and industry

Test every design and production change in simulation, 13x faster than one by one.
A paper with its claims highlighted, each one linked to the reference behind it

Academic research

Rerun analyses across every paper and dataset, 3.5x faster.
Price candles, green for gains, and red for losses, with predicted candles in violet

Finance and strategy

Test a thesis against every comparable company 3.5x faster, with a source on each figure.
A sales funnel narrowing from all accounts through researched and qualified accounts to pipeline, with accounts dropping out along the way

Go-to-market

Research every account in a market, 3.5x faster.
A processor running kernel variants, where a few variants run faster than in the last round

Compute and hardware

Benchmark every kernel, code change, and hardware variant 13x faster than one by one, and map the supply chains behind them.
Our Mission

We want everyone to have the tools for research and discovery.

Read Our Mission
A crowd of people of different ages, each with their own agent. Some are asking a question and others already have a result.

Run your research on Antinuous.

Start with our agent, or connect Claude, ChatGPT, or your own agent to the same compute, orchestration, simulation, and artifacts.

Sign up

How we measured

Speedups compare production runs from September to November 2026 with the same work done one job at a time. Agent swarms of 20 or more agents ran 3.5x faster at the median, and the largest, 253 agents, ran 19x faster. Compute jobs on 16 or more workers ran 13x faster at the median, and the largest turned 3 hours of sequential runs into 8 minutes. Half of research runs finish within 4 minutes.