Autonomous Company

Autonomous Company (formerly Auto-Company) is an open-source framework by Zheyuan (Max) Kong and contributors that runs a team of AI agents in a continuous loop on a local machine. It defines 14 agent roles, each modeled on a well-known expert's approach, and works within human-configured goals, permissions and budgets. It runs on macOS and on Windows through WSL, using Claude Code by default or Codex CLI.

What It Is

The project is a multi-agent orchestration layer around command-line LLM engines. A loop script reads a prompt and a shared work summary (memories/consensus.md), calls the chosen engine, forms a team of roles as needed, executes research, coding, deployment or marketing work, updates the summary and waits before the next cycle. Each cycle is an independent CLI call. Product files, logs, configuration and usage data persist between cycles. Deployment and publication depend on the tools and authorization available.

Team Roles and Skills

The 14 roles span strategy, product, engineering, business and intelligence layers, for example CEO, CTO, QA, DevOps/SRE, Marketing, CFO and Research Analyst. The team skill suggests picking 2-5 relevant roles per task. More than 30 reusable skills, such as deep research, scraping, financial modeling, SEO, security audit and UX audit, can be loaded on demand. The default workflow guidance moves from brainstorming to validation to a go/no-go decision and then implementation.

Dashboard and Product Center

A local web dashboard shows runtime status, usage, budgets, logs and per-cycle reports. The Product Center catalogs multiple product runs, with read-only import of existing runs, explicit work requests, a configurable limit on parallel projects (four by default), batch creation and templates. A catalog of 22 runnable example products is included.

Operation and Safeguards

On macOS the loop runs as a launchd daemon, and on WSL as a systemd --user service, with auto-restart. It includes a circuit breaker, rate-limit backoff and consensus recovery after failed cycles. Humans steer by editing Next Action or Human Overrides in the consensus file while the run is stopped. The charter gives agents behavioral rules such as not force-pushing or deleting repositories, but these are prompt rules and not an OS sandbox.

Current Status

The README describes the project as experimental. Each cycle consumes model quota, and default engine permissions allow broad local actions. Optional Cursor and OpenAI-compatible adapters require explicit configuration. The latest listed release is v3.0.0.