Pre-flight Checklist
Problem or Motivation
AG-UI's Dojo showcases agent frameworks, but none demonstrate live multi-agent orchestration over real OS processes with human-in-the-loop interrupts on genuine CLI permission prompts.
CLI Agent Orchestrator (CAO) runs each coding agent (Kiro CLI, Claude Code, Codex, …) as a real process in an isolated tmux session and coordinates them supervisor→worker over MCP. It already exposes an AG-UI-compliant surface (see awslabs/cli-agent-orchestrator#458 — the AG-UI Phase 2 L2 construct library + run plane). What's missing is a first-class Dojo integration so users can see a real multi-agent fleet — including a genuine interrupt/approval round-trip — rendered through AG-UI.
Proposed Solution
Add cli-agent-orchestrator as a first-class AG-UI integration:
- A TypeScript thin client (
@ag-ui/cli-agent-orchestrator) — a HttpAgent subclass pointing at CAO's AG-UI endpoint (zero extra wire code).
- A keyless example server (via CAO's
mock_cli fleet — no API keys) exposing the four standard features: agentic_chat, shared_state, human_in_the_loop, and the flagship interrupt (a real permission prompt → RUN_FINISHED outcome={type:"interrupt"} → resume[]).
- Dojo wiring (menu/agents/env), content-json, four Playwright e2e specs, and a CI matrix entry — all keyless so the e2e suite runs without secrets.
Because the fleet is real OS processes, the interrupt feature demonstrates something no API-wrapper integration can: approving/denying a live agent's actual permission prompt.
Alternatives Considered
- An API-wrapper integration — rejected: it loses the real-CLI fidelity (native tool behavior, auth, provider features) that is CAO's entire value proposition.
- Hosting a live multi-agent demo server — deferred to a maintainer-side follow-up (tmux/Docker on the hosted Dojo).
Additional Context
Pre-flight Checklist
Problem or Motivation
AG-UI's Dojo showcases agent frameworks, but none demonstrate live multi-agent orchestration over real OS processes with human-in-the-loop interrupts on genuine CLI permission prompts.
CLI Agent Orchestrator (CAO) runs each coding agent (Kiro CLI, Claude Code, Codex, …) as a real process in an isolated tmux session and coordinates them supervisor→worker over MCP. It already exposes an AG-UI-compliant surface (see awslabs/cli-agent-orchestrator#458 — the AG-UI Phase 2 L2 construct library + run plane). What's missing is a first-class Dojo integration so users can see a real multi-agent fleet — including a genuine interrupt/approval round-trip — rendered through AG-UI.
Proposed Solution
Add
cli-agent-orchestratoras a first-class AG-UI integration:@ag-ui/cli-agent-orchestrator) — aHttpAgentsubclass pointing at CAO's AG-UI endpoint (zero extra wire code).mock_clifleet — no API keys) exposing the four standard features:agentic_chat,shared_state,human_in_the_loop, and the flagshipinterrupt(a real permission prompt →RUN_FINISHED outcome={type:"interrupt"}→resume[]).Because the fleet is real OS processes, the interrupt feature demonstrates something no API-wrapper integration can: approving/denying a live agent's actual permission prompt.
Alternatives Considered
Additional Context
cli-agent-orchestrator[agui]extra), so there is no cross-repo drift risk.