AI has left the chat window. Now it needs coordination.
Fleets of specialized agents are spreading across clouds, factory floors, vehicles, and devices — pinned to the data, machines, and infrastructure that won't move.
The model isn't the bottleneck anymore. Making distributed agents behave like one system is.
Three forces are pulling AI out of a single cloud and into a distributed system.
01
Multi-provider is the default.
Anthropic, OpenAI, Google. Bedrock, Foundry, Vercel, Cloudflare. No model, cloud, or agent runtime will coordinate across its competitors. Winning systems will mix all of them.
02
Some artifacts can't move.
The most valuable data is the data that can't leave its environment. Agents need to live where the data is, not the other way around.
03
AI will drive the most value where it interacts with the physical world.
Factories, vehicles, devices: the intelligence has to come to them.
Distributed agents have to discover each other, address each other, and coordinate across providers they don't control. Request and response, stitched together with retries, is not a foundation for that. It's a workaround.
HTTP · workaround
NATS · foundation
03 — A new role
Behind every agent orchestrator is a meta agent.
A new role is emerging: the agent orchestrator. A person whose expertise is monitoring fleets of agents and course-correcting in flight.
Behind every orchestrator will be a meta agent: the programmatic tool that reduces context switching and codifies their workflow so they can manage to outcomes, not manage agent work.
Back to Unix: small composable primitives.
We're not shipping the meta agent. We're shipping the fabric every meta agent will run on.
Swap any layer. Build the meta agent that fits your business.
01Connectivity that spans cloud to factory.
02Identity for every agent type and every running instance.
03Durable state and handoff built in.
04An audit trail of every interaction, on by default.
