
TL;DR: We’re building models and RL environments to train agents to work swarm efficiently, benchmarks to measure their performance, and infrastructure to run them at scale.
video link: https://www.youtube.com/watch?v=kBTQD1hAuyE
We’re building the foundations of agent swarms that can materially accelerate all work: benchmarks to measure model capabilities, models trained to coordinate thousands of agents, RL environments to improve them, and infrastructure that can put a datacenter’s worth of compute behind a single conversation within seconds.
We both graduated top 10% from Cambridge. Ram is a published ML researcher and Varun worked at the frontier of parallel programming languages and shipped enterprise multi-agent systems at Palantir and Tracelight. We realised that agents were good enough to solve most use cases. So why couldn’t we run thousands of them at once and finish hours of work in minutes?
We started by building open-source infrastructure capable of running 10,000 agents in a single session. But scaling the number of agents exposed a deeper problem: today’s models struggle to coordinate, and swarms can perform worse than a single agent.
This led us to build SwarmBench, an independent benchmark measuring how well models delegate tasks to other agents and ability to speculate i.e. anticipating upcoming work and starting agents on it early. Our evaluations isolate these coordination bottlenecks across different swarm architectures and model flavours. Can thousands of Opus 5.5 agents working together outperform a single instance? Our benchmarks shows that more isn’t always better: performance depends on the type of workload (not necessarily scale) where coordination failures can consistently erase the benefits of scale.
Check out initial public findings now acyclic.dev/benchmarks
Finally, we’re building swarm coordination RL environments and have begun to train models to work better together. Our benchmarks expose gaps in how models delegate tasks and explore competing approaches in parallel. We turn those failures into training environments, then benchmark again to measure improvement. We’re building a continuous loop for improving swarm coordination: measure, train, repeat.
If you’re training models, building with agents, or investing in the future of AI, we’d love to hear from you: ram@acyclic.dev.