"We replaced four coordination services with one runtime call. The mesh just figures it out."
A runtime for swarms of agents to coordinate, share state, and reach consensus without a single point of control.
Three composable primitives replace the scheduler, the broker, and the coordination layer in a typical agent stack. Everything else is a topology choice.
Agents discover peers and negotiate work directly. No central scheduler, no bottleneck. Topology emerges from capability.
Agreement primitives that tolerate partitions, recover cleanly, and never deadlock. Vote, gather, decide — in three lines.
Same primitives in datacenter, on a laptop, or on a single-board device. One topology, every machine.
Conflict-free replicated stores keep every agent's view eventually identical, even across continents and intermittent links.
Model Context Protocol is native. Tools, memories, and capabilities are addressable — every agent speaks the same dialect.
Distributed traces, peer-level metrics, and a topology view that updates in real time. The mesh is never a black box.
Discover capable peers. Run the work. Reach consensus on the result. The runtime exposes nothing else — every higher-order pattern is a few lines of glue on top.
// route a task across the mesh import { atara } from "@atara/runtime"; const peers = await atara.discover({ capability: "summarize.long", region: "us-east, eu-west", minPeers: 3 }); const result = await atara.run("agent.summarize", { peers, quorum: 2, timeout: 8_000 }); const consensus = await atara.vote(result.candidates);
Every agent in the global mesh routes for itself. Watch traffic flow across regions in real time, and trust that nothing in the middle can fail.
"We replaced four coordination services with one runtime call. The mesh just figures it out."
Technical preview is invite-only. We'll review your application within a week.