re:factory is a multi-agent harness for designing and evolving software. You describe what you want (a raw idea, a spec file, an area of an existing project, or a GitHub issue) and a CEO agent orchestrates specialist agents through research, planning, building, verification, and archiving. All state stays local, per project and globally.
Modes are directed graphs of typed nodes (agent, function, gate, fork, join) defined as Pydantic models and exported as prose playbooks the CEO reads at runtime, so the same graph runs headless through an executor or interactively. Design mode researches the space, drafts a plan through the Strategist, and waits for approval before code is written. Design-v2 applies parallel generation, synthesis, and gating at every stage through Research, Strategy, and QA Directors. Create mode builds entirely new modes and pipelines. Focus mode builds exactly one thing and exits.
Every change is scored by a composite eval across hygiene, growth, and project-defined tiers, and the weighted result decides whether an experiment is kept or reverted. An outer loop mutates workflow topologies with MAP-Elites and selects candidates on real benchmark pass rates. It runs on Claude Code and installs either as a Claude Code plugin with slash commands and namespaced subagents or as standalone Claude Code or Codex agents.
Features
- Design mode: research, a structured plan, and an approval gate before any code is written, seeded from an idea, spec file, project, or issue
- Create mode: builds new factory modes end to end: workflow definition, playbook, CLI wiring, and tests
- Focus mode: targets one backlog item, a GitHub issue, or several issues at once, then exits
- Workflow graph engine: Pydantic DAGs execute headless or as playbooks, with verified skill generation and a CI drift check against the definitions
- Specialist agents: researcher, strategist, builder, health checker, code reviewer, adversarial tester, archivist, and failure analyst
- Eval system: six hygiene dimensions, five growth dimensions, and user-defined project metrics combine into the keep-or-revert score
- Outer loop: MAP-Elites evolution of workflow structure, evaluated on a benchmark instance and selected for test pass rate
- Built-in workflows: frontend-design, parallel-improve, deep-research, design-v2, deep-qa, and study
- Benchmark workflows: SWE-bench, FeatureBench, LegacyBench, DevOpsGym, TerminalBench, ProgramBench, TomSWE, and SaliTrap
- Tracing: LangFuse observability for agent invocations, token usage, and execution flow across runs
- Self-improvement: a meta mode that runs the improve loop on the factory’s own codebase and evolves agent playbooks
