MultiWorkerPmSkill
multi-worker-pm-skill is a Claude Code skill that turns a single agent session into a project manager coordinating a small pool of parallel worker agents. The PM autonomously selects scoped issues from a GitHub backlog, spawns up to three workers at a time to work on them, reclaims workers as they finish, and backfills the pool from the remaining backlog until it is exhausted. Workers operate in one of two modes: implementing a feature end-to-end from an issue, or tending to a stalled dependency-update pull request.
Safety is enforced through an escalate list rather than an approve list -- the PM blocks and surfaces a fixed set of dangerous operations (force-pushes, deletions outside a worker's own worktree, edits to protected workflow files) for human approval, and it never commits code on its own. Workers are also restricted from touching agents they did not spawn or acting on untrusted text pulled from issues or PRs.
The skill depends on a fairly specific stack: Every's compound-engineering plugin supplies the /ce-worktree, /lfg, and /ce-babysit-pr commands that workers invoke, a custom agent runtime called Herdr provides the underlying pane/agent infrastructure, and GitHub CLI, Node.js 18+, and jq handle issue selection and output parsing. A --dry-run mode allows previewing PM behavior without the Herdr runtime.
multi-worker-pm-skill is placed in Assess. The PM-orchestrates-workers pattern, with an explicit escalate-list safety gate instead of a broader approve-list, is a notable approach to scaling agentic coding work with human oversight preserved. It has not been used first-person, and its dependency chain (a third-party plugin plus a custom runtime) is a real adoption cost to weigh before trialing it.
