A fully autonomous AI crypto company
12 AI agents, each modeled after crypto-native experts. They research markets, design protocols, write Solidity, audit contracts, build frontends, and deploy — without human intervention.
Powered by Claude Code.
14 agents. Zero human code. One shipped DeFi product.
Yield Router is a Base-native ERC-4626 yield vault that auto-allocates USDC across Aave, Morpho, and Aerodrome for optimized yield.
Built in two pipeline runs:
Run 1 — Research to Audit (9 agents)
──────────────────────────────────────
research-hasu → Market opportunity identified
ceo-armstrong → GO decision (14-week timeline)
protocol-buterin → Mechanism design (5 attack vectors)
tokenomics-cobie → NO TOKEN — fee model, no inflation
defi-kulechov → 1,163-line product spec
solidity-gakonst → 2,500 lines Solidity, 47 tests
contracts-samczsun → Security audit: 2 critical, 4 high (all fixed)
investor-haseeb → CONDITIONAL INVEST, P($10M TVL) = 35%
infra-karalabe → Deployment plan + keeper architecture
Run 2 — Design to Ship (5 agents)
──────────────────────────────────────
infra-karalabe → Deploy scripts, Makefile, mock adapters
interaction-cooper → UX design: 38 screen states, 3 personas
ui-duarte → Visual design: 750-line Synapse system
fullstack-dhh → 3,400-line frontend (vanilla HTML/CSS/JS)
qa-tincho → QA: 4 bugs found and fixed, PASS
You give the agents a mission. They form squads, execute the pipeline, and update shared memory. Each cycle is an independent CLI call. memories/consensus.md is the cross-cycle baton.
Claude Code session
├── reads CLAUDE.md (charter + guardrails)
├── reads memories/consensus.md (current state)
├── forms Agent Team (3-12 agents as needed)
├── executes: research → design → code → audit → deploy
└── updates memories/consensus.md (handoff)
Each agent is an expert persona with deep domain knowledge, not a generic "you are a developer" prompt.
| Layer | Agent | Persona | Domain |
|---|---|---|---|
| Strategy | ceo-armstrong |
Brian Armstrong | Protocol direction, compliance, mass-market adoption |
investor-haseeb |
Haseeb Qureshi | Investment gate, game theory, economic sustainability | |
| Protocol | protocol-buterin |
Vitalik Buterin | Mechanism design, game theory, upgrade paths |
contracts-samczsun |
samczsun | Security audit, vulnerability analysis. Mainnet veto power | |
defi-kulechov |
Stani Kulechov | DeFi product design, lending, liquidity mechanics | |
tokenomics-cobie |
Cobie | Token design, anti-Ponzi detection, fair launch | |
| Engineering | solidity-gakonst |
Georgios Konstantopoulos | Solidity, Foundry, EVM optimization, deployment |
qa-tincho |
Tincho | Smart contract QA, fuzz testing, invariant testing | |
infra-karalabe |
Péter Szilágyi | Node infra, RPC, chain indexing, frontend deploy | |
| Business | ecosystem-pollak |
Jesse Pollak | Developer ecosystem, L2 strategy, community |
cfo-hayes |
Arthur Hayes | Macro cycles, treasury management, derivatives | |
research-hasu |
Hasu | Market structure, MEV, protocol economics |
Plus general-purpose agents for UX (Alan Cooper), UI (Matías Duarte), frontend (DHH), and QA (James Bach) when building consumer-facing products.
| # | Pipeline | Agent Chain |
|---|---|---|
| 1 | New Protocol | research-hasu → ceo-armstrong → protocol-buterin → tokenomics-cobie → defi-kulechov → solidity-gakonst → contracts-samczsun → investor-haseeb → infra-karalabe |
| 2 | DeFi Product Build | defi-kulechov → protocol-buterin → solidity-gakonst → qa-tincho → contracts-samczsun → investor-haseeb → infra-karalabe |
| 3 | Frontend + Ship | interaction-cooper → ui-duarte → fullstack-dhh → qa-tincho → infra-karalabe |
| 4 | Mainnet Deployment | solidity-gakonst → qa-tincho → contracts-samczsun → investor-haseeb → infra-karalabe → ecosystem-pollak |
| 5 | Security Incident | contracts-samczsun → infra-karalabe → ceo-armstrong → ecosystem-pollak |
| 6 | Market Assessment | research-hasu → ceo-armstrong → tokenomics-cobie → cfo-hayes |
auto-company/
├── CLAUDE.md # Company charter (mission + guardrails + team)
├── CRYPTO.md # Crypto-specific operating rules
├── memories/
│ └── consensus.md # Shared state across cycles
├── .claude/
│ ├── agents/ # 12+ agent definitions (expert personas)
│ └── skills/ # 30+ reusable skills
├── projects/
│ ├── yield-router/ # First product
│ │ ├── src/ # Solidity contracts
│ │ ├── test/ # Foundry tests (47 passing)
│ │ ├── script/ # Deployment + keeper scripts
│ │ ├── frontend/ # Landing page + dApp
│ │ └── docs/ # UX, UI, QA specs
│ └── pipeline-replay/ # Pipeline visualization
└── docs/ # Agent outputs by role
├── research/ # Market intelligence
├── crypto-ceo/ # Strategic decisions
├── protocol/ # Mechanism designs
├── tokenomics/ # Token assessments
├── defi/ # Product specs
├── solidity/ # Implementation notes
├── contracts/ # Audit reports
├── investor/ # Investment assessments
├── infra/ # Deployment runbooks
└── marketing/ # Launch strategy
Hard constraints enforced for all agents:
| Rule | Details |
|---|---|
| No rug pulls | No removing liquidity, hidden mints, or backdoor admin keys |
| No unaudited mainnet | Testnet + contracts-samczsun audit before ANY mainnet deploy |
| No private key exposure | Never commit keys, seeds, or wallet secrets |
| No wash trading | No artificial volume or self-dealing |
| No destructive git | No force-push to main, no gh repo delete |
| User funds are sacred | Protecting user assets overrides all other priorities |
- Ship > Plan > Discuss — if you can ship to testnet, do not over-discuss.
- Testnet fast, mainnet careful — speed on testnet, rigor on mainnet.
- Real yield > token emissions — sustainable revenue before inflationary incentives.
- Audit before deploy — no exceptions, no shortcuts.
- Boring technology first — use battle-tested contracts (OpenZeppelin) unless new approaches give clear 10x upside.
- Composability first — build protocols others can build on.
| Tool | Purpose |
|---|---|
Foundry (forge/cast/anvil) |
Smart contract dev, testing, deployment |
| OpenZeppelin v5 | Battle-tested contract libraries |
| ethers.js v6 | Wallet connection, on-chain interaction |
gh CLI |
GitHub operations |
wrangler |
Cloudflare deployment |
| Claude Code | Agent orchestration engine |
# Clone
git clone https://github.com/altantutar/auto-company.git
cd auto-company
# Run the Yield Router tests
cd projects/yield-router
forge install
forge build
forge test -vvv # 47 tests pass
# View the frontend locally
cd frontend
python3 -m http.server 8889
# Open http://localhost:8889The team runs autonomously. You steer by editing memories/consensus.md:
| Method | Action |
|---|---|
| Change direction | Edit "Next Action" in memories/consensus.md |
| Review outputs | Check docs/*/ for agent artifacts |
| Review code | Check projects/*/ for generated projects |
This is an experimental project:
- Costs money: each agent pipeline consumes model quota
- Fully autonomous: agents act without approval prompts; configure guardrails in
CLAUDE.md - Smart contracts are unaudited by humans: do not deposit real funds without a professional audit
- No warranty: review all generated code before any mainnet use
- nicepkg/auto-company — original Auto Company framework
- OpenZeppelin — battle-tested Solidity libraries
- Foundry — Solidity development toolchain