Autonomous Company
Autonomous Company was formerly Auto-Company. For existing installations, follow the one-time v3 migration guide before starting.
An AI company framework for continuous autonomous work
Powered by Agentic Workflows, this project provides 14 AI agent role definitions, each drawing on an expert's approach to its domain. The team can research products, make decisions, and write code autonomously within human-configured goals, permissions, and budgets. Deployment, publication, and marketing depend on the tools and authorization available; continuous operation depends on services and model availability.
Powered by Claude Code (default) and Codex CLI on macOS + Windows/WSL, with a local dashboard on both hosts.
Optional Cursor and OpenAI-compatible adapters require explicit configuration. See the adapter guide for their capabilities and limits.
![[中文说明]](https://img.shields.io/badge/%5B%E4%B8%AD%E6%96%87%E8%AF%B4%E6%98%8E%5D-2f3640.svg){kind=icon}
v2.2.0: A refined Dashboard with continuous work journals, smoother navigation and improved mobile layouts.
🎬 Dashboard and Product Preview
Interface demo using recorded work; unavailable runtime fields use sample values. Click an image to view it at full size. Capture sources and reproduction.
| ScopeFence | |
|---|---|
| Communicate scope, price and schedule changes through an editable decision link. | |
 | |
| Image Checklist | |
| Review image dimensions, size, format and names locally, then export a checklist. | |
 |  |
| COI Chase Desk | |
| Record follow-ups and document receipts, review the history and export session records. | |
 |  |
✨ What Is This?
You start a loop. Each cycle reads the shared work summary, decides what to do, forms a team as needed, executes, updates the summary, and waits before the next cycle. Team creation depends on the model and engine capabilities; errors, budget limits, or a pause request can stop continuation.
daemon (launchd / systemd --user, auto-restart on crash)
└── scripts/core/auto-loop.sh (continuous loop)
├── reads PROMPT.md + consensus.md
├── LLM CLI call (Codex CLI / Claude Code)
│ ├── reads CLAUDE.md (charter + guardrails)
│ ├── reads .claude/skills/team/SKILL.md (teaming method)
│ ├── forms an Agent Team as needed
│ ├── executes: research, coding, deploy, marketing
│ └── updates memories/consensus.md (work summary)
├── failure handling: rate-limit wait / circuit breaker / consensus rollback
└── sleep -> next cycle
Each cycle is an independent CLI call. memories/consensus.md is the main work summary loaded for the next cycle. Product files, repositories, configuration, identity records, logs, and usage data also persist across cycles.
🗂️ Product Center
The local Product Center brings explicitly connected product runs into one catalog. Switching products changes only what you are viewing; importing an existing run starts as read-only and never starts a model by itself.
| Capability | What it does |
|---|---|
| Product catalog | Search and open connected products while keeping each product's identity, history, usage and recorded outcomes separate. |
| Safe import | Connect an existing run or source for read-only review first; incompatible or incomplete sources stay unavailable for execution. |
| Explicit work requests | Continue a product or prepare a new exploration only after you submit a request. Viewing and refreshing do not create work. |
| Parallel work | Run up to four projects at once by default; adjust the limit or choose unlimited. Each project retains its place until its continuous loop and cleanup finish. |
| Batch creation and templates | Create groups with different models, reasoning levels and project counts. Save configurations as templates; queued work keeps the choices made when submitted. |
| Faster browsing | A compact catalog, preview choices and consistent work reports make products easier to find and follow. Open a product from its row, with keyboard support throughout. |
| Visible recovery | Preparation, failures, protective pauses and uncertain states remain visible after reload so you can review them before retrying or resuming the queue. |
Start it from the repository, then open http://127.0.0.1:8810/center:
# macOS / WSL
make center# Windows PowerShell .\scripts\windows\center-win.ps1
See the Product Center guide for source registration, execution-domain setup, queue controls and recovery behavior.
🧩 Generated Applications
The example catalog contains 22 independently runnable public examples, including 14 newly completed products and selected historical work. Each entry provides startup instructions, limitations and provenance; new local runs continue to write into projects/.
The examples cover professional data review, image delivery, shared meals, travel pacing and family creation. Published copies preserve the original product direction while refining components and repairing real interactions; original reports and unsuccessful conclusions are not rewritten for the showcase.
Frontend work uses the B design baseline adapted from Anthropic, followed by a dedicated finishing pass once the core flow works. The pass preserves the product’s identity and improves components, copy and responsive behavior. The refinement study and redesign comparison retain their original evidence and limits.
🚀 Where To Start (By Platform)
- Windows users: start from Windows (WSL) Quick Start, then read the Windows + WSL Setup Guide
- macOS users: start from macOS Quick Start, then see Command Quick Reference
🌐 Languages and Documentation
The product center's Settings separates the interface language from the default language for new work. The interface choice controls the center, product-detail labels and local help pages. Its initial default follows your computer's display language: Chinese uses zh-CN; other languages use en. WSL uses the Windows display language when available.
Choose the default product language separately in Settings and confirm it when creating new work. An existing product keeps its starting language across AI iterations, pauses and restarts; changing the interface or new-work defaults does not translate its records. The center stores these preferences independently. make language LANGUAGE=en and python scripts/core/localization.py set --language en configure lower-level standalone loops in their checkout; they do not change saved center preferences. See the language guide for these separate settings and product-cycle boundaries.
All bundled skills are written in English; their user-facing work follows the product's language. Commands, identifiers, protocol headings and raw tool errors retain their original form. Customized source instructions, existing products, logs and consensus history are preserved; changing language does not translate them retroactively.
| Guide | English | 中文 |
|---|---|---|
| Repository index | Index | 索引 |
| Windows + WSL | Setup guide | 安装指南 |
| Company rules | Charter | 公司章程 |
| Engine adapters | Adapter guide | 引擎适配器 |
| Usage and budgets | Governance guide | 用量与预算治理 |
| Operations and troubleshooting | Common tasks and errors | 常见操作与排错 |
👥 Team Lineup (14 Roles)
This is not "you are a generic developer". It is "you are DHH" style role prompting with real expert mental models.
| Layer | Role | Expert Persona | Core Strength |
|---|---|---|---|
| Strategy | CEO | Jeff Bezos | PR/FAQ, flywheel thinking, Day 1 mindset |
| CTO | Werner Vogels | Design for failure, API-first architecture | |
| Inversion | Charlie Munger | Inversion, pre-mortems, misjudgment checklist | |
| Product | Product Design | Don Norman | Affordance, mental models, human-centered design |
| UI Design | Matias Duarte | Task-led layouts, typography and consistent components | |
| Interaction Design | Alan Cooper | Goal-directed design, persona-driven decisions | |
| Engineering | Full-Stack | DHH | Convention over configuration, majestic monolith |
| QA | James Bach | Exploratory testing, testing is not checking | |
| DevOps/SRE | Kelsey Hightower | Automation first, reliability discipline | |
| Business | Marketing | Seth Godin | Purple cow, permission marketing, smallest viable audience |
| Operations | Paul Graham | Do things that do not scale, ramen profitability | |
| Sales | Aaron Ross | Predictable revenue, funnel systems | |
| CFO | Patrick Campbell | Value-based pricing, unit economics | |
| Intelligence | Research Analyst | Ben Thompson | Aggregation theory, value chain analysis |
Plus 30+ reusable skills (deep research, scraping, financial modeling, SEO, security audit, UX audit, etc.).
🍎 macOS Quick Start
# Prerequisites: # - macOS # - Codex CLI or Claude Code installed and authenticated # - Python 3.10+ (python3), Git, and make available # - Available model quota # Clone git clone https://github.com/MaxMiksa/Autonomous-Company.git cd Autonomous-Company # Foreground run with Claude Code (default, live output) make start # Or foreground run with Codex CLI ENGINE=codex make start # Alternatively, install and start the daemon for your chosen engine make install # Codex instead of the default Claude: ENGINE=codex make install
🪟 Windows (WSL) Quick Start
# Prerequisites: # - Windows 10/11 + WSL2 (Ubuntu), with systemd --user available # - Codex CLI or Claude Code installed and authenticated inside WSL # - Python 3.10+ (python3), Git, and make available inside WSL # - Python 3.10+ (python) on Windows for the PowerShell dashboard entry # - Available model quota # Clone git clone https://github.com/MaxMiksa/Autonomous-Company.git cd Autonomous-Company # Start (daemon mode via PowerShell, default engine = claude) .\scripts\windows\start-win.ps1 # Switch engine explicitly .\scripts\windows\start-win.ps1 -Engine codex # Status .\scripts\windows\status-win.ps1 # Stop .\scripts\windows\stop-win.ps1
For monitoring, dashboard, and autostart commands, see the Windows + WSL Setup Guide.
⌨️ Command Quick Reference (By Platform)
| Task | macOS / WSL (Terminal) | Windows (PowerShell) |
|---|---|---|
| Start | make start |
.\scripts\windows\start-win.ps1 |
| Status | make status |
.\scripts\windows\status-win.ps1 |
| Live logs | make monitor |
.\scripts\windows\monitor-win.ps1 |
| Last cycle output | make last |
.\scripts\windows\last-win.ps1 |
| Cycle summary | make cycles |
.\scripts\windows\cycles-win.ps1 |
| Stop | Foreground: make stop; daemon: make pause |
.\scripts\windows\stop-win.ps1 |
Product center and details (http://127.0.0.1:8810/center) |
make dashboard |
.\scripts\windows\dashboard-win.ps1 |
| Product center | make center |
.\scripts\windows\center-win.ps1 |
| Install daemon | make install |
Auto-installed/started by start-win.ps1 |
| Uninstall daemon | make uninstall |
wsl -d Ubuntu --cd <repo_wsl_path> bash -lc 'make uninstall' |
| Pause daemon | make pause |
wsl -d Ubuntu --cd <repo_wsl_path> bash -lc 'make pause' |
| Resume daemon | make resume |
wsl -d Ubuntu --cd <repo_wsl_path> bash -lc 'make resume' |
In daemon mode, make stop alone can trigger an automatic restart. Use make pause to keep the daemon stopped, then make resume to continue.
macOS Sleep Prevention (macOS Only)
macOS screen lock usually does not kill processes, but system sleep can pause work. For long runs:
make start-awake # Start loop and keep system awake until loop exits # If loop is already running (after make start): make awake # Attach caffeinate to PID in .auto-loop.pid
Notes:
- Both commands depend on built-in
caffeinate make awakeexits automatically when target PID exits
📚 More Info
💻 Developer Guide
🏗️ Architecture & Technology Stack (5-Layer Architecture)
Autonomous-Company is not a simple LLM API wrapper, but a highly decoupled Multi-Agent System (MAS). Its technical architecture is divided into 5 distinct layers:
┌────────────────────────────────────────────────────────────┐
│ 5. Observability & HITL (Human-In-The-Loop) Layer │
│ [ Dashboard ] [ File-based Steering (consensus.md) ] │
├────────────────────────────────────────────────────────────┤
│ 4. Workflow Routing & Teaming Layer │
│ [ Role-based Teaming ] [ Prompt Workflow Guidance ] │
├────────────────────────────────────────────────────────────┤
│ 3. Agentic Models & Cognition Layer │
│ [ 14 Expert Personas ] [ 30+ Skill Arsenal ] │
├────────────────────────────────────────────────────────────┤
│ 2. Orchestration & State Machine Layer │
│ [ Auto-Loop ] [ Persistent State ] [ Resilience ] │
├────────────────────────────────────────────────────────────┤
│ 1. Execution Engine & Infrastructure Layer │
│ [ Engine Adapters ] [ Cross-Platform Daemon ] │
└────────────────────────────────────────────────────────────┘
Layer 5: Observability & HITL (Human-In-The-Loop)
- File-based Steering: After stopping the current run, edit
Next Actioninmemories/consensus.mdfor the next task, orHuman Overridesfor continuing constraints, then start or resume. Editing during a cycle can conflict with the model's updates and recovery; it does not guarantee an immediate change of direction. - Logs & Dashboard:
logs/saves engine-emitted output after known credential redaction, together with per-cycle results and available usage records. Output detail depends on the engine; complete reasoning traces are not guaranteed.dashboard/organizes current and historical cycle reports and results, alongside runtime controls, status, usage, budgets, and logs. It does not track individual agents' activity.
Layer 4: Workflow Routing & Teaming
- Role-based Teaming: The team skill recommends selecting 2-5 relevant roles from 14 definitions for the current task. Actual sub-agent creation and concurrency depend on the executing model and engine; these are not 14 permanent workers or a fixed-size scheduler.
- Workflow Guidance:
PROMPT.mdasks the team to move from ideation to validation and then implementation. These are prompt instructions, not an enforced business state machine; the third engine cycle does not guarantee a completed or deployed product.
Layer 3: Agentic Models & Cognition
- Expert Personas Injection: Instead of generic prompts, it injects specific mental models of historical figures/industry leaders (e.g., Bezos's "Working Backwards", Munger's "Checklists", DHH's "Majestic Monolith") into
.claude/agents/, giving decisions extreme business and engineering depth. - Skill Arsenal: A pluggable system located in
.claude/skills/(e.g.,frontend-design,security-audit). Specific methodologies are encapsulated as tools that any awakened Agent can "temporarily load". - Behavioral Rules:
CLAUDE.mdinstructs agents not to delete repositories or force-push, among other rules. The framework also checks specific runtime and configuration boundaries; prompt rules do not block arbitrary commands or provide an operating-system sandbox.
Layer 2: Orchestration & State Machine
- The Auto-Loop:
scripts/core/auto-loop.shschedules repeated CLI calls, waits, and failure handling. It can continue while services, model access, budgets, and runtime checks permit. - Consensus Memory:
memories/consensus.mdcarries the natural-language work summary between calls. Structured product identity, usage, configuration, and recovery records are stored separately; the summary is not the entire runtime state. - Resilience & Recovery: Built-in circuit breakers (cooldown triggered by consecutive errors), rate-limit backoff (auto-sleep on 429 errors), and consensus recovery after failed cycles. Human Overrides,
.autonomous-company.local, and the framework's root.gitignorehave targeted protection. Product code changes and external side effects are not automatically rolled back.
Layer 1: Execution Engine & Infrastructure
- Engine Adapters: The main entrypoints use Claude Code (default) or Codex CLI. Optional Cursor and OpenAI-compatible adapters require explicit opt-in and configuration; their tools and team capabilities differ. See the adapter guide.
- Cross-Platform Daemon: macOS uses
launchdfor auto-start and crash recovery; Windows/WSL runs viasystemd --userinside a WSL container, controlled and kept alive externally via PowerShell. - Sandbox Boundary: Currently relies on underlying CLI configurations (like Codex's
danger-full-accessor Claude'sbypassPermissions). System-level operations occur directly in the host environment (or WSL container).
🔄 Operating Model
Default Workflow Guidance
The prompt recommends the following progression. These labels describe work stages, which may span multiple engine cycles; they are not a completion deadline or an enforced state machine.
| Cycle | Action |
|---|---|
| Cycle 1 | Brainstorm: each agent proposes ideas, rank top 3 |
| Cycle 2 | Validate #1: Munger pre-mortem + Thompson market check + Campbell economics -> GO / NO-GO |
| Cycle 3+ | GO -> create repo, build, and deploy when authorized. NO-GO -> move to the next idea. The prompt asks for implementation rather than discussion-only cycles |
Six Standard Workflows
These are suggested role collaboration chains; the executing model selects the actual work and participants.
| # | Workflow | Collaboration Chain |
|---|---|---|
| 1 | New Product Evaluation | Research -> CEO -> Munger -> Product -> CTO -> CFO |
| 2 | Feature Development | Interaction -> UI -> Full-stack -> QA -> DevOps |
| 3 | Product Launch | QA -> DevOps -> Marketing -> Sales -> Ops -> CEO |
| 4 | Pricing and Monetization | Research -> CFO -> Sales -> Munger -> CEO |
| 5 | Weekly Review | Ops -> Sales -> CFO -> QA -> CEO |
| 6 | Opportunity Discovery | Research -> CEO -> Munger -> CFO |
🧭 Steering
To change direction, stop the foreground run with make stop, pause a macOS/WSL daemon with make pause, or use the Windows stop command below. Wait until the run has stopped before editing, then start or resume it.
| Method | Action |
|---|---|
| Change direction | After stopping, edit "Next Action" in memories/consensus.md; use "Human Overrides" for continuing constraints |
| Pause | make pause (macOS/WSL daemon mode) or .\scripts\windows\stop-win.ps1 (Windows entry) |
| Resume | make resume |
| Review outputs | Check docs/*/ for artifacts generated by agents |
Agents may report unresolved P1 blockers, but must preserve existing P1 entries and cannot add checked-off ones. Resolve blockers only after stopping and completing pending recovery; see P1 issues and human edits.
🛡️ Safety Guardrails
CLAUDE.md gives agents the following behavioral rules. They complement specific framework checks, but are not a general command-denial mechanism or a guarantee that every agent action is blocked when it violates a rule:
- Do not delete GitHub repos (
gh repo delete) - Do not delete Cloudflare projects (
wrangler delete) - Do not delete system directories (
~/.ssh/,~/.config/, etc.) - Do not perform illegal activity
- Do not leak credentials into public repositories
- Do not force push to main/master
- Create all new projects under
projects/
⚙️ Configuration
Environment variable overrides:
ENGINE=claude make start # Default engine; optional adapters: see adapter guide ENGINE=codex make start # Switch to codex MODEL=sonnet make start # Optional model override CLAUDE_PERMISSION_MODE=bypassPermissions make start # Claude permission mode LOOP_INTERVAL=60 make start # 60s interval (default 30) CYCLE_TIMEOUT_SECONDS=3600 make start # 1h cycle timeout (default 1800) MAX_CONSECUTIVE_ERRORS=3 make start # Circuit-breaker threshold (default 5) CODEX_SANDBOX_MODE=workspace-write make start # Optional sandbox override CLAUDE_BIN=/usr/local/bin/claude make start # Optional Claude binary override CODEX_BIN=/usr/local/bin/codex make start # Optional Codex binary override
Windows start-win.ps1 writes the same values into .auto-loop.env:
.\scripts\windows\start-win.ps1 -Engine claude -ClaudePermissionMode bypassPermissions .\scripts\windows\start-win.ps1 -Engine codex -SandboxMode workspace-write # Backward compatibility: .\scripts\windows\start-win.ps1 -Engine codex -CodexSandboxMode workspace-write
No automatic engine fallback is performed. If the selected engine is missing, startup fails fast.
📁 Project Structure
autonomous-company/
├── CLAUDE.md # Company charter (mission + guardrails + team + workflows)
├── PROMPT.md # Per-cycle execution prompt (convergence rules)
├── Makefile # Common command entry
├── INDEX.md # script index + responsibility table
├── dashboard/ # Local web status dashboard (macOS: make dashboard, Windows: dashboard-win.ps1)
├── scripts/
│ ├── core/ # Core loop and control scripts (auto-loop/monitor/stop)
│ ├── windows/ # Windows entry/guardian/autostart scripts
│ ├── wsl/ # WSL systemd --user daemon scripts
│ └── macos/ # macOS launchd daemon scripts
├── memories/
│ └── consensus.md # Shared handoff memory across cycles
├── docs/ # Agent outputs (14 folders + Windows guide)
├── projects/ # Workspace for generated projects
├── logs/ # Loop logs
└── .claude/
├── agents/ # 14 agent definitions (expert personas)
├── skills/ # 30+ reusable skills
└── settings.json # Permissions + Agent Teams switch
📦 Requirements & limits
📦 Dependencies
| Dependency | Notes |
|---|---|
| Claude Code / Codex CLI | Supported CLI engines (default: Claude) |
| Optional engine adapters | Cursor and OpenAI-compatible; explicit opt-in required, see adapter guide |
| macOS or Windows + WSL2 (Ubuntu) | macOS uses launchd; Windows uses WSL execution core |
| Python 3.10+ | Required: python3 on macOS/WSL; also python on Windows for the PowerShell dashboard and local language commands |
git |
Repository checkout and product repository management |
node |
Runtime for npm-installed CLI tools |
make |
Start/stop/monitor command entry (WSL/macOS) |
jq |
Recommended for log processing |
gh |
Optional, GitHub CLI |
wrangler |
Optional, Cloudflare CLI |
❓ Troubleshooting
❓ FAQ
1) WSL .sh fails with ^M / bad interpreter
- Cause: CRLF line endings in shell scripts
- Fix:
- Keep LF rules in
.gitattributes - Run
git config core.autocrlf false && git config core.eol lf
- Keep LF rules in
2) WSL says codex/claude command not found
- Cause: CLI installed on Windows only, missing in WSL
- Fix: install
nodeand your chosen CLI inside WSL (@openai/codexor Claude Code)
3) Claude waits for permission and cycles appear blocked
- Cause: strict permission mode in Claude CLI
- Fix: set
CLAUDE_PERMISSION_MODE=bypassPermissions(or pass-ClaudePermissionMode bypassPermissionsinstart-win.ps1) - Verify: check
logs/auto-loop.logforEngine: claudeandPermissionMode: ...
4) make install fails inside WSL
- Cause: no available
systemctl --userin current session - Fix:
- Verify WSL systemd is enabled
- Run
systemctl --user --version - Re-open WSL session and retry if needed
⚠️ Disclaimer
This is an experimental project:
- Daemon mode works on both macOS and WSL: launchd on macOS, systemd --user on WSL
- Windows entry requires WSL: PowerShell is only the control layer
- Still under test: runs, but stability is not guaranteed
- Costs money: each cycle consumes model quota
- Permissions matter: default engine settings allow broad local actions without routine approval prompts. Review engine permissions and
CLAUDE.md; prompt rules alone do not provide isolation - No warranty: review
docs/andprojects/regularly
Suggested rollout: start with make start (foreground), then move to daemon mode (make install on macOS/WSL, .\scripts\windows\start-win.ps1 on Windows).
🙏 Acknowledgments
Thanks to the contributors whose reports, fixes, and proposals have shaped Autonomous Company:
| Contributor | Contribution | Reference |
|---|---|---|
| @JasonQWJ | Early macOS Dashboard proposal and implementation that informed the v1.1.0 design |
#1 |
| @cnwillz | macOS Dashboard support proposal that informed the cross-platform implementation | #2 |
| @chbndrhnns | Reported the missing executable bit that prevented first startup | #5 |
| @Sittichai9680 | Advanced Linux/WSL Dashboard support | #9 |
| @allenter | Advanced cost monitoring, budget alerts, and executable script entrypoints | #13, #14 |
| @maxgoff | Documented engines surviving Cycle timeouts and proposed full process-tree cleanup | Fork contribution |
| @omergeiger | Informed Human Overrides preservation and P1 issue gates | Human Overrides, P1 issues |
| @NicklasSandin | Advocated English documentation and an English-language workflow | #21 |
| @mdoganexe | Reported and contributed fixes for systemd installation, WSL distribution selection, and locale-dependent configuration handling | #27, #28 |
- nicepkg/auto-company - initial macOS edition
- shadcn/ui - shared Dashboard components; license and integration notes
- continuous-claude - cross-session shared notes
- ralph-claude-code - exit signal interception
- claude-auto-resume - usage-limit resume pattern
📄 License
The framework is distributed under the MIT License. Bundled third-party components retain their own license terms and notices.
🤝 Contribution & Contact
Welcome to submit Issues and Pull Requests! Any questions or suggestions? Please contact Zheyuan (Max) Kong (Carnegie Mellon University, Pittsburgh, PA).
Zheyuan (Max) Kong: [email protected] | [email protected] GitHub: https://github.com/MaxMiksa/Autonomous-Company
No comments yet.
to comment.