long-running-agent

To convert a planning spec into an autonomous agent workflow, this Skill reads your SPEC.md file and translates it into a structured phase-based task setup. It generates a claude-task.json and claude-prompt.md in your project root for multi-session execution.

Yes, an AI agent can run autonomously across multiple phases using the orchestrator mode. This mode allows the agent to operate continuously across phase boundaries without requiring human verification at each step.

To set up phase-based tasks with human verification checkpoints, use the phase-checkpoint mode. This mode decomposes your spec into structured phases, pausing the autonomous workflow for explicit human verification at each phase boundary before proceeding.

To generate a long-running agent from a planning document, you need a SPEC.md file in your repository. The workflow uses this input to output a claude-task.json, a claude-prompt.md, and the preserved SPEC.md in your project root.

The best way to structure multi-session execution for an autonomous agent is using a phase-based task structure with standardized templates. This approach decomposes plans into distinct phases and tasks equipped with machine-checkable verifications.

You should use a long-running agent for phase-based work when your project requires multi-session execution with explicit verification checkpoints. It is designed for complex specs that need autonomous operation across multiple phases rather than single-prompt tasks.