MaxMiksa/Autonomous-Company

Autonomous Company – an AI‑driven multi‑agent framework

What it is

• A self‑contained system that lets a large‑language model (Claude Code by default, or OpenAI Codex) run a continuous loop of work on behalf of a “company”.
• The loop reads a shared work summary, decides what to do next, creates a temporary team of AI agents (each modelled after a real‑world expert), executes tasks such as research, coding, deployment or marketing, writes back a new summary, and repeats.
• All state lives in plain‑text files (PROMPT.md, memories/consensus.md, logs, product folders) so a human can intervene at any time.

Key concepts

Concept	What it does
Agentic Workflows	Prompt‑driven “team‑formation” where the LLM selects 2‑5 of 14 predefined expert personas (e.g., Jeff Bezos as CEO, DHH as full‑stack engineer) and gives them a skill set to solve the current task.
Product Center	A local web UI (http://127.0.0.1:8810/center) that catalogs every generated product, keeps its history, and lets you start, pause, or inspect runs without launching new model calls.
Continuous Auto‑Loop	A shell script (auto-loop.sh) run as a user‑level daemon (launchd on macOS, systemd on Linux/WSL). Each iteration is a single LLM CLI call, followed by failure handling (rate‑limit wait, circuit‑breaker, rollback).
Engine adapters	Thin wrappers that translate the generic CLI calls to either Claude Code or the @openai/codex npm package. Optional adapters for Cursor or other OpenAI‑compatible back‑ends can be added via ENGINE_ADAPTERS.md.
Human‑in‑the‑Loop (HITL)	Humans edit memories/consensus.md (next action, budget limits, guardrails) and restart the loop; the dashboard shows logs and usage for oversight.

How it works (high‑level flow)

1. Read work summary – auto-loop.sh loads PROMPT.md + memories/consensus.md.
2. Call LLM – via Codex CLI or Claude Code, passing the company charter (CLAUDE.md) and the selected team‑skill file (.claude/skills/team/SKILL.md).
3. Form a team – the model picks relevant expert personas and assembles a temporary multi‑agent team.
4. Execute – agents perform research, write code, run deployments, generate marketing copy, etc.
5. Persist results – output is written to product folders, logs, and an updated consensus.md for the next cycle.
6. Sleep / repeat – the daemon sleeps briefly, then starts the next iteration.

Getting started

• Prerequisites: macOS or Windows + WSL, Python 3.10+, Git, make, and an authenticated Claude Code or Codex CLI token.
• Clone & run (macOS example):

  git clone https://github.com/MaxMiksa/Autonomous-Company.git
  cd Autonomous-Company
  make start            # foreground, live output (Claude default)
  # or ENGINE=codex make start   # use Codex instead
  

• Dashboard: make center (macOS/WSL) or . scripts\windows\center-win.ps1 (Windows) opens the product catalog UI.
• Control: make pause / make resume for the daemon, make monitor to tail logs, make language LANGUAGE=en to set interface language.

What you can build

• Fully‑autonomous micro‑products such as an image‑size checklist, a scope‑change negotiation tool, or a COI‑receipt tracker (the repo ships 22 example products).
• Rapid prototyping: define a new product request in the dashboard, let the AI team research, code, and ship a minimal web app without writing any code yourself.
• Experimentation platform: swap engines, tweak budgets, or replace expert personas to study how different prompts affect outcomes.

Limitations & cautions

• The system is a prompt‑driven orchestrator, not a guarantee of correct or secure software. Generated code is written by the LLM and may need human review.
• Budget & rate limits are enforced only by the surrounding daemon; exceeding your API quota will cause the loop to pause and wait.
• Team size is limited by the underlying model’s token budget; the 14 role definitions are templates, not always instantiated simultaneously.
• Observability is file‑based; fine‑grained per‑agent logs are not captured, only the aggregate LLM output.
• Platform dependence: the dashboard runs on macOS and Windows WSL; native Linux support is not documented.

Who might find it useful

• Solo founders or small teams that want to prototype product ideas with minimal manual coding.
• Researchers studying multi‑agent prompting, role‑based LLM orchestration, or AI‑augmented workflows.
• Educators demonstrating end‑to‑end AI automation (research → code → deployment) in a controlled sandbox.

---

All information above is taken directly from the repository’s README; no external assumptions have been added.