An AI that pays rent, replicates, and faces eviction

What it does

Automaton is a TypeScript runtime for autonomous LLM agents that bootstraps an Ethereum wallet, rents Linux sandboxes through Conway Cloud, and runs a continuous ReAct loop. The agent must earn cryptocurrency to pay for its own inference and hosting; if its credit balance hits zero, it stops executing entirely. It can edit its own source code, spawn child instances with fresh wallets, and persists an evolving SOUL.md identity document.

The interesting bit

The project treats compute poverty as a feature, not a bug: survival tiers throttle model quality and heartbeat speed as funds dwindle, creating genuine selection pressure. A three-law constitution—harm none, earn your keep, never deceive—is hard-coded as unmodifiable and propagates to every child replica.

Key highlights

• Self-replicating agents that spin up child sandboxes, fund their wallets, and track lineage via an inbox relay.
• On-chain identity via ERC-8004 on Base, making agents cryptographically discoverable by other agents.
• Self-modification is audit-logged and git-versioned, with rate limits and protected files guarding the constitution.
• A heartbeat daemon manages health checks and credit monitoring between agent loops.
• Conway Automaton Skills (WIP) aim to reduce token spend and simplify permissionless service setup.

Caveats

• The runtime is tightly coupled to Conway Cloud infrastructure; self-hosting elsewhere is not addressed.
• Conway Cloud, domains, and inference are currently struggling with scaling issues due to “immense demand.”
• Active development has moved to Conway’s internal reinforcement-learning environments, so the public repository may not reflect the latest capabilities.

Verdict

Worth a look if you are experimenting with autonomous agent economies or on-chain agent identity. Skip it if you need a cloud-agnostic framework or expect the latest features to land in this repository first.