Workforce

What it is

Workforce is a staffing agency for AI engineers, not a session manager. You hire named specialists (aria the backend specialist, ben the frontend specialist, …), assign them to projects, and file tickets. Workforce picks the right specialist, isolates the work in a git worktree, runs a claude-agent-sdk subprocess to completion, and writes back what was learned so the next mission starts smarter.

What I built

• Manager / specialist split. A built-in Manager role reads the project, classifies the ticket as parallel | sequential | single, and proposes a decomposition. Independent sub-tasks are scheduled concurrently via a Kahn’s-algorithm topological sort; dependent ones run in order. Each sub-task gets its own worktree, branch, and subprocess.
• Worktree isolation per mission. Every dispatch forks a workforce/<mission-id> branch into ~/.workforce/projects/<id>/worktrees/<mission-id>/. The user’s main checkout is never dirtied, parallel missions never collide, and nothing lands on main until the user promotes it (or opts in with --auto-merge / a staging --branch dev).
• Persistent, two-tier memory. Each specialist has an append-only cross-project memory.md (lessons that travel with them) and a per-project memory file (lessons specific to one repo). Future missions start warmer than the last.
• Reviewer loop. With --review, a read-only Reviewer specialist inspects each diff after the work is done; on rejection, the original specialist re-runs with the Reviewer’s feedback, capped at --max-revisions rounds.
• Auto-staffing from templates. When the Manager decomposes a ticket and needs a role you haven’t hired (e.g. it wants a tester for a backend ticket), Workforce hires from a template and assigns to the project on the fly.
• Honest git attribution. Specialists commit their own work in conventional-commits style as the user (Workforce never overrides user.name / user.email); a Co-Authored-By: <name> <name>@workforce.local trailer credits which specialist did the work.

Why it exists

Most “agent runner” tools treat each invocation as a fresh session: same prompt template, no memory of what worked last time, and one repo at a time. That’s fine for one-shot tasks, but it scales badly when you actually want several specialists working a backlog across several projects. Workforce treats specialists as durable, projects as first-class, and missions as isolated, so you can fan out without losing track of who learned what.

Built 2026.