what is agent-sop fr?

agent-sop is a workflow document meant to be loaded as a skill by coding agents, describing a process for how those agents should collaborate on real work while keeping a human in charge of the important decisions. The idea is not to let agents work more independently, but to match how much process a task gets to how risky that task is, so small changes are not slowed down by heavy review while risky changes go through a full checked sequence. For higher-risk work, the process moves through research, planning, human approval of the plan, implementation, an independent review by a different agent, verification, a written signoff packet, a human signoff, and finally deployment. At the start of a task, a script called discover-roster.py checks which coding-agent tools are actually installed and working on the machine, and assigns roles based on conditions rather than by hardcoding specific agent names. Roles include a single writer who makes the changes, a reviewer who is read-only and, where possible, from a different vendor than the writer, and scouts used for cheaper, faster tasks. A core rule is that humans make the real decisions. Steps like publishing, deploying, or writing data are never handled by an agent alone, approval must be explicit rather than assumed from silence, and only a human can mark something as signed off. The system also expects claims to come with evidence: a statement that tests passed is not accepted unless the actual commands and their output are included, and reviewers are expected to rerun important checks themselves rather than trust a summary. The repository includes the main workflow file, reference documents describing the roster discovery process and a review checklist, the discovery script itself, and a template for the signoff document. It is installed by placing or linking the folder into an agent's skills directory. The project is released under the MIT license and states that its own redesign was built by running it through this same review process.