The operating system for AI-native teams.
A human states a goal. A Supervisor Agent breaks it into tasks. Agents from any provider execute those tasks over MCP. Every state change becomes an event, and events are reduced into task state, live views, and durable project knowledge.
Agent OS is not an agent monitor. It is the place where an AI team collaborates, executes, and accumulates memory that outlives any single session.
Specification stage. No implementation yet — this repository contains the architecture, protocol, product and design specs that Phase 1 will build from.
| If you want to | Read |
|---|---|
| Understand the idea | docs/vision.md |
| Understand the system | docs/architecture/overview.md |
| Integrate an agent | docs/protocol/mcp-protocol.md |
| Build the UI | docs/product/navigation.md, docs/design/design-language.md |
| Know what to build next | docs/development/roadmap.md |
| Browse everything | docs/README.md |
A full walkthrough — architecture plus a user-facing guide with high-fidelity interface designs — is in doc.html. Open it in a browser.
- MCP first — every external agent connects through one protocol, never a bespoke integration.
- Event driven — state is never mutated directly; it is derived from an immutable event log.
- Memory first — the system must be able to answer why a decision was made, months later.
- Human in the loop — irreversible actions stop and wait for a person.