AgentPlane

About

Teams adopting coding agents need an enforceable record of who did what, not chat transcripts. AgentPlane divides responsibility three ways: humans set outcomes and approve material risk, coding agents do semantic work inside bounded episodes, and the CLI (semantically blind, mechanically authoritative) owns task state, Git/PR routing, and evidence. Each task advance emits a packet with an objective, writable scope, context, and a typed result schema; agents cannot perform lifecycle transitions or claim formal approvals. Workflows run in direct mode for solo reversible work or branch_pr mode with worktrees, branches, and hosted checks. The tool is aimed at engineering teams that want Git itself to be the durable review surface for agent-driven changes.