We built Axon because deploying an AI agent should feel like deploying a web app.

Not a research experiment. Not a proof-of-concept. A real, production agent with tools, memory, stop conditions, and a managed runtime — live in the cloud from a single command.

The problem with existing frameworks

Most agent frameworks are DIY runtimes in disguise. You wire up your own loop. You manage context windows manually. You implement your own stop conditions, tool dispatch, and state persistence. By the time you have something production-grade, you've spent more time on infrastructure than on the agent itself.

That's not building an agent. That's building a runtime that happens to call an LLM.

What Axon does differently

Axon separates the two concerns that every agent framework conflates: what the agent does, and how cognition works.

You write what only you can write:

  • What the agent knows — its identity, its boot prompt, its domain
  • What it can do — tools, server routes, plugins
  • How it behaves — policy, scope, access controls

Everything else — the cognitive loop, context assembly, compression, tool dispatch, stop conditions, memory — is handled by Cognos, our proprietary cognition runtime. You never touch it. You never need to understand it.

const { stream } = axon.stream({ prompt: [session, task] })

Everything past that call is Axon's concern.

An agent is a folder

my-agent/
├── axon.config.ts    # identity, engine, policy
├── src/              # boot prompt, tools, prompts, scripts
├── server/           # HTTP routes and plugins
├── data/             # durable storage and knowledge
└── modules/          # installed capabilities

That folder is the agent. Commit it. Deploy it. Roll it back. It runs on your laptop in dev, headlessly in CI, and as a live cloud service — same source, zero code changes.

axon deploy

axon deploy

That's it. Axon packages your agent, provisions managed compute on Google Cloud Run, attaches a persistent GCS bucket, and returns a WebSocket endpoint. $10/month. Fully managed. No infrastructure to think about.

What's next

We're working on the module registry — pre-built tools, prompts, and server routes you can install into any agent with a single command. Think npm for agent capabilities.

The engine system is coming next: swap engine: "axon-astra-v1" in your config to upgrade your agent's cognition with no code changes.

We're early. The foundation is solid. Come build something.