AI agents can already do real work. But they do it one step at a time, so a complex task still takes hours or days. [3]
Running many agents at once sounds simple. In practice, someone has to isolate each agent, divide the work and merge the results without conflicts. [3]
Acyclic Labs builds that infrastructure.
Hand a task to a swarm and Acyclic fans it out across many agents, each in its own fork of the workspace. They work in parallel, then merge into one patch or pull request. [1][3]
Instead of one agent working through a long task step by step, a swarm finishes it in a fraction of the time. [3]
Fan out, fork, verify, merge
Start from the command line in your project. Sign in, describe the task, such as fixing a flaky test, and watch the agents work live. Then pull the merged patch into your checkout or open a pull request. [1]
You can also start a swarm from the dashboard or connect a coding agent over MCP. [1]
For bigger jobs, Acyclic’s managed platform fans the work out across many agents, forks the ones worth exploring, shows every branch as it runs and keeps only the results you verified. A single swarm session can scale to thousands of cores. [1][4]
The building blocks for running agents in parallel are open source. [3]
Today, teams use Acyclic to run heavy document-processing pipelines in minutes instead of hours. [3]
Two Cambridge researchers who built agents for enterprises
Ram Vinjamuri and Varun Latthe met on their first day studying computer science at the University of Cambridge. [2]
Vinjamuri’s undergraduate research on transfer learning for Vedic Sanskrit beat the state of the art and was published at ACL. Latthe wrote a dissertation on an auto-parallelising compiler. [2][3]
Then both built multi-agent systems in production. Vinjamuri was a forward deployed engineer at Palantir, shipping systems for clinical trial design for pharma clients. Latthe led agent architecture at Tracelight for Big Four accounting firms. [3][4]
The agents were good enough to solve most problems, they realized. So why couldn’t they run thousands at once and make everything faster and cheaper? Acyclic is their answer. [4]
Acyclic is based in London and is part of Y Combinator’s Fall 2026 batch. [3]
How a team would use it
Here’s how two kinds of teams would use it.
A flaky test in a big codebase
An engineering team hands Acyclic a flaky test. Agents investigate different parts of the failure at once, each in its own sandbox, and the team pulls back one merged patch to review and apply. [1]
A heavy document pipeline
A team processing a large batch of documents splits the work across a swarm. Each agent takes its share in its own fork of the workspace, and the results merge into one output, in minutes instead of hours. [3]
Acyclic is onboarding teams gradually, starting with those that already run agents in volume: startups with domain-specific agents, companies with agent swarms in production, and teams running many Claude Code or Codex sessions on large codebases. [1]
Where Acyclic Labs ranks
Most agents work through a task one step at a time. Acyclic splits it across a swarm and merges the results. [3]
The bigger picture
Acyclic’s founders believe the next big gains will come from how agents are run, not only from the models. [3]
Their thesis is that agentic work should happen in real time, however complex the task. People should spend their time on ideas and design, while the building happens at the speed of thought. [4]
Acyclic is betting that the way to get there is many agents at once, not one agent working faster.
Too big for one agent is no longer too big.
Written by SOTA2 from Acyclic Labs’s public materials. Customer experiences and product plans are reported by Acyclic Labs. Ranking analysis and comparison data are from SOTA2.
