Skip to main content
Impala Swarm is a way to run an agent swarm on Impala: a parent agent that fans out into many subagents over a shared context, with the loop itself executing on Impala rather than in your application.

​
Why run the swarm here

Prefix reuse across siblings. Subagents spawned from one parent share a large prompt prefix — system prompt, tool schemas, the parent’s accumulated context. When the loop runs next to the fleet, that KV state is already resident rather than being re-sent and recomputed per child. This is the mechanism described in Impala Herd, applied to fan-out. Swarm state survives capacity changes. A long-running swarm outlives the machines it starts on. Impala holds the execution state — the parent context, the subagents in flight, what has already completed — so a run continues across scaling events and instance changes in your cloud instead of starting over. Cost per completed task, not per call. A swarm is the clearest case of the argument in Run async and open source: nobody is reading the intermediate tokens, dozens of children are in flight, and the only number that matters is what the finished task cost.

​
Where it runs

Swarm runs in Bring your own cloud deployments, so the swarm’s state and every intermediate result stay inside your own VPC and buckets — see Security & data handling.

​
Need help?

Reach out to your Impala contact directly.