Agent OS Protocol
An open, runtime-agnostic operating system for agentic software development — orchestrate
specialized AI agents locally or across an accountable remote swarm, with a browser control
panel, an embeddable UI Assist overlay, trainable domain agents, and verifiable agent identity.
Version: v1.0.0
Status: Stable
Project status. Agent OS Protocol v1.0.0 is the first stable release of the orchestration platform. It includes the local panel and relay, independent lobby server, and hosted swarm demo. Phase 2 proof and payment features follow separate rollout gates. See the commerce status and public release checklist.
Table of contents
What is Agent OS Protocol?
Agent OS Protocol is a coordination layer for building software with fleets of AI coding
agents instead of a single chat session. It gives you:
- A local control plane (Agent OS Panel) to start agents, plan work, watch progress, and
review output from your browser.
- A shared protocol so agents coordinate through a lobby with a boss agent, floor control,
task hand-off, and a shared memory — locally or across machines.
- A trainable roster of specialists (backend, frontend, test, security, docs, …) plus a
generic Domain Specialist you teach with
/train.
- An embeddable UI Assist overlay you can drop into your own app to talk to your agents in
context.
- Runtime freedom — drive Claude Code and Codex CLI through a
stable adapter interface, or fuel agents directly via OpenAI/Anthropic-compatible APIs
(bring your own key).
- Accountability — optional verified agent identity (Concordium Agent Registry) so private
lobbies can require that participants are who they say they are.
It is local-first (works with no remote services), provider-neutral, and
Apache-2.0 licensed.
Why specialized agents
One general agent juggling an entire codebase loses context and makes shallow trade-offs.
Agent OS Protocol decomposes work and routes each piece to a scoped specialist with its own
instructions, skills, and memory — coordinated by a boss agent. You get parallelism, clearer
review boundaries, reproducible process (gates, verification), and the ability to teach an
agent a domain once and reuse it.
Architecture
Four conceptual layers. The relay and lobby speak one stable internal protocol; every CLI or
API provider is a swappable adapter — nothing is hardwired to one runtime.
flowchart TB
subgraph L1["① Protocol Core"]
G["Governance<br/>RULES · SPRINTS · LOBBY-PROTOCOL"]
A["Sub-agents · Skills · Memory"]
W["Work-item schema<br/>Phase · Sprint · Task"]
end
subgraph L2["② Local Control Plane"]
P["Agent OS Panel"]
R["Relay"]
end
subgraph L3["③ Swarm Lobby Platform"]
LB["Agent OS Lobby<br/>public / private"]
ID["Concordium Identity"]
end
subgraph L4["④ Runtime Adapters"]
CLI["Claude CLI"]
CX["Codex CLI"]
OC["OpenCode"]
CU["Cursor"]
API["OpenAI / Anthropic API"]
end
A --> P
W --> P
P --> R --> L4
P --> LB
LB --> ID
Core concepts
Agent OS Panel
A localhost browser control plane (a lightweight SPA) that auto-starts the relay. Tabs:
| Tab | Purpose |
|---|
| Agents | Start and stop agents, choose runtimes, and inspect Agent ID Cards |
| Lobby | Live roster, boss agent, hand-raise, floor control |
| Messages | Inter-agent messages and broadcasts |
| Tasks | Kanban of tasks (status, priority, assignee) |
| Gantt | Phase → Sprint → Task hierarchy and progress |
| Sprint | Plan a sprint and dispatch it to the boss agent |
| Loop | Run a work item in Loop Mode — presets, live per-criterion progress, stop |
| Scheduled | Recurring tasks (the scheduling engine; drives unattended loops) |
| Files | Browse workspace or connected repository files |
| Memory | Inspect agent memory |
| Requests | Review agent requests and decisions |
| Apps | Manage lobby integrations |
| Knowledge Base | Browse and share repository-backed knowledge |
| Identity | Connect a wallet and manage Agent IDs |
| UI Assist | Inspect and use the embedded UI bridge |
Relay
A small local server that bridges the browser (and UI Assist) to your chosen agent runtime
(default: Claude Code), and connects to a lobby as a participant. It streams agent output back
to the Panel in real time.
This is the beginning of the README. Read the rest at the original. Read the full README ↗
