CAR (codex-autorunner) provides low-opinion agent coordination tools for automating complex, long-running AI agent implementations. It allows users to define plans, convert them into CAR-compatible tickets, and delegate execution to agents without constant supervision, leveraging familiar tools like Git and Python. This system amplifies agent capabilities, enabling users to focus on higher-level tasks.

How It Works

CAR functions as a state machine processing user- or agent-generated tickets. Each agent receives its ticket, workspace context, and prior outputs. The system treats the filesystem as the data plane and tickets as the control plane, relying on agents for accurate instruction execution. This design avoids constraining AI models, using them as the execution layer.

Quick Start & Requirements

The fastest setup involves passing the setup guide to a preferred AI agent for an interactive, environment-specific walkthrough. Alternatively, from a source checkout, ./car --help bootstraps a local Python virtual environment and installs CAR if dependencies are missing. No specific hardware (e.g., GPU, CUDA) or Python version requirements are detailed.

Highlighted Details

• Multi-interface Control: Offers a Web UI for setup/chat/tickets, a Telegram interface for on-the-go interaction, and a Project Manager Agent (PMA) for delegating CAR operations to AI.
• Collaborative Workspace: Uses the filesystem as a shared scratchpad for context and artifact exchange, enabling collaborative editing of markdown tickets and workspace documents by users and agents.
• Agent Agnosticism: Designed for easy integration with any Agent Client Protocol (ACP) compatible agent.
• Mobile Responsiveness: The Web UI is mobile-friendly.

Maintenance & Community

The README provides no details on maintainers, community channels (Discord/Slack), sponsorships, or a public roadmap. Users interested in supporting new agents are encouraged to reach out or open a pull request.

Licensing & Compatibility

The provided README content does not specify the software license, preventing an assessment of compatibility for commercial use or closed-source linking.

Limitations & Caveats

CAR's performance depends on the AI model's capability; weaker models may yield suboptimal results. The system is unsuitable for agents prone to "scope creep" or "reward hacking." The Telegram interface has fewer features than the Web UI, and the Web UI's authentication token option is noted as not extensively battle-tested.