What is deputy-agent?

Deputy is a TypeScript framework for running long, unattended AI tasks without constant human involvement. You give it a task description, it generates a custom plan and process tailored to that task, then runs the work through a structured sequence of stages, checking in with you only at key decision points. The design uses a master and worker arrangement rather than having multiple AI agents collaborate as equals. A master role called Meta does the planning, sets up the task-specific process, and judges outcomes. A worker role does the actual execution. Two additional roles observe the worker: a Watcher monitors output in real time, and a Reviewer gives verdicts at stage gates. This four-role structure is meant to catch errors and correct the work without needing a human to supervise every step. The framework is aimed at general knowledge work rather than coding tasks specifically. The README notes it is built for jobs that run roughly one hour to two days, not for quick interactive exchanges. Each role can use a different AI model or provider, so you can assign a cheaper model to the worker and a more capable one to the Meta role. Both Claude and Codex are supported. Agents coordinate through files written to disk, so the worker's state persists across multiple sessions. You interact with Deputy through either a local web interface or a command-line tool. The web interface shows live progress, conversation streams, and agent output in one place. The CLI exposes the same operations and is suited to scripting. This is version 0.1.0 and described as a reference implementation derived from a higher-level design specification that is not included in the repository. The author notes that behavior is tuned primarily for Claude, that Codex and GPT models perform less well, and that cross-task memory is not yet implemented. It is not described as production-ready. The project is licensed under Apache 2.0.