SOP — Agent Collaboration Workflow

First principle — Agent-first. Every skill we build or publish, this one included, is operated by agents, not hand-configured by humans. Anyone using a skill has at least one agent, so the skill must guide that agent to do its own onboarding: detect the environment, ask the human only for decisions (never for mechanical setup), and write any required configuration, roster, or paths itself. A skill step that requires a human to edit files by hand is a design bug. Humans own decisions, approvals, and signoffs — agents own everything else.

Use this skill as the local operating procedure. For high-risk work, it provides the full auditable sequence:

research → plan → human approval → implementation → independent review → verification → signoff packet → human signoff → deploy

Lower-risk work uses the lighter workflow in §1.1. The skill is deliberately conservative. It is not permission to change production services, expose ports, restart shared units, or modify private data without an explicit user decision.

0. Capability slots and roster discovery

• Human: owns scope, accepts trade-offs, signs documents, and approves destructive or operational actions.
• orchestrator: classifies the task, assigns slots, enforces gates, and synthesizes artifacts and reports.
• sole-writer: is the only agent allowed to edit the declared repository, worktree, or production configuration during a phase.
• reviewer: uses an independent read-only session to test requirements and evidence; it never approves its own writing.
• scout: performs bounded read-only exploration, research, or monitoring and returns evidence.
• monitor: tracks long-running background work (≥2 min) via artifact-file checks, pane status, and logs; identifies stalls and reports progress to the human. Usually carried by the orchestrator; see §5.1.