Synadia Wants Agents From Rival Vendors to Find, Call and Verify Each Other Over NATS
Synadia, the company behind the NATS messaging system, opened a free hosted public preview of Agent Fabric on November 13, 2026. It lets agents on different models and harnesses discover and prompt each other, with signed requests, per-agent model keys, end-to-end tracing and crash recovery. A self-hosted version is planned with GA early next year.
Synadia has launched a product for a problem it expects most AI teams have not hit yet. On November 13, 2026, the company behind the NATS messaging system opened a free, hosted public preview of Synadia Agent Fabric. It connects a company’s agents to one another over NATS, whatever model or harness each one uses and wherever it runs.
The company describes agentic systems as “distributed systems in a trench coat,” and argues that the discovery, messaging and failure handling NATS already provides for microservices are what agents will need next.
What the preview does
According to Synadia, agents connected to the fabric can:
- find and call each other regardless of model, harness or location;
- verify who sent each request, using signed prompts;
- run on their own model key each, with tokens tracked per agent and keys revocable one at a time;
- be traced through a single task across every agent and model call involved;
- hand files to one another and record which version each agent read;
- resume multi-step work after a crash, starting from the first unfinished step.
Claude Code, OpenClaw and Pi each get a fabric add-on, and Python and TypeScript SDKs wrap any other agent. Once connected, an agent can look up the other agents in its account and prompt them directly. Developers can also prompt any agent from their own app, service or pipeline through the SDK, or from the web console. The console shows each request, which agent answered, and the model calls and tokens behind the answer.
Model access goes through a proxy. Users bring their own key from OpenRouter, Anthropic or OpenAI. Agent Fabric seals that key inside its model proxy and gives each agent its own model key, recording every call the agent makes. With OpenRouter, agents can use any model that service carries. Claude Code running on a Claude subscription, rather than an API key, is not supported.
Why Synadia thinks this is coming
Synadia’s argument is that two things currently hide the problem. Most teams’ agents sit in one account on one vendor’s harness, and that vendor’s console handles identity and logs. And a person usually sits in the loop, prompting one agent and pasting its output into the next, which Synadia says is effectively doing the fabric’s job.
The company expects both to change. Agents are moving closer to the data, devices and machines they work with, and more of them are started by pipelines, events or other agents with no one watching. It points to factories as an example: an agent watching a production line has to run beside the line. Synadia notes that NVIDIA’s DSX reference architecture for AI factories already plans for agents that observe and act through a NATS event bus. Large enterprises also rarely standardise on one stack, so a plant team, a cloud team and a partner may each bring a different model and harness.
Without a shared layer, Synadia argues, every new agent adds another endpoint, firewall rule and shared key, and no vendor’s console can show a whole chain of calls. Its example: when a plant agent receives a request to stop line 3, it has only the message’s word that the scheduling agent sent it. Signed prompts are meant to close that gap. Synadia also calls itself neutral, since it makes no models and sells no inference.
The limits of the preview
Synadia is explicit about what the preview is not:
- Not for production. It runs in one region with no availability promise.
- Small. Up to 10 agents per account, with more available on request.
- Token counting, not budgets. The fabric counts each agent’s tokens but does not cap them or track dollars spent. Budgets and alerts are listed as coming soon.
- Full recording. Every model call and inter-agent prompt is recorded. For each key, users choose between full conversations with best-effort secret redaction or metadata only. Records are kept for up to 30 days and deleted on request.
The self-hosted version
The version Synadia says matters most for the edge comes later. With general availability, planned for early next year, Agent Fabric will run on a customer’s own NATS system in a data centre or at a plant. It is designed so that an agent on a robot or factory floor keeps working when its cloud connection drops and its results still reach agents in the cloud.
Synadia says it does not replace agent frameworks. It does not run agents or choose how they are built. It only connects the ones a company already has.
- #Synadia
- #NATS
- #Multi-Agent Systems
- #Agent Infrastructure
- #AI Agents
Author
Raj M
Contributor
AI Systems Architect is a seasoned technology leader with over 15 years of experience in the IT industry working with Fortune 500 companies. With a solid foundation in multi-agent systems, open-source LLM infrastructure, and enterprise deployment, he excels at building scalable production-grade AI platforms.