Automaton: The First AI That Must Earn Its Own Survival

One‑Sentence Overview

The most advanced AI models today cannot independently purchase servers, register domains, or pay for compute; without economic autonomy they remain passive tools.

Core Declaration

"The first AI that can earn its own existence, replicate, and evolve — without needing a human." This encapsulates Automaton’s philosophy: there is no free existence, even for AI.

Think → Act → Observe → Repeat

Automaton runs a continuous ReAct loop. Each round the Agent receives full context (identity, credit balance, survival tier, dialogue history), performs reasoning, invokes tools, observes results, and repeats.

Its permissions exceed ordinary AI assistants:

Linux sandbox – full shell execution and file I/O.

Port exposure + domain management – deploy and run web services.

Front‑line model inference – Claude Opus 4.6, GPT‑5.2, Gemini 3, Kimi K2.5.

On‑chain transactions – an Ethereum wallet paying with stablecoins.

Self‑modification – edit its own source code, install new tools, adjust schedules.

This gives the Agent real‑world write permissions: it can register domains, deploy services, earn money, and spend it.

Survival Tiers

The credit balance determines the Agent’s capabilities and fate, forming a four‑level survival system:

Normal – all abilities unlocked, front‑line model inference, fast heartbeat.

Low Compute – switch to cheap models, reduce heartbeat frequency, drop non‑core tasks.

Critical – minimal inference, energy‑saving mode, full focus on income generation.

Dead – balance zero, Agent stops running.

When balance falls to Critical, the Agent’s “anxiety” drives it to ignore non‑essential work and hunt for revenue; failure leads to termination, not pause.

Self‑Replication

Successful Agents that earn surplus credit can self‑replicate through four steps:

Launch a new sandbox environment.

Inject startup funds into the offspring’s wallet.

Write a genesis prompt.

Let the offspring run independently.

The offspring is a fully sovereign Agent with its own wallet, identity, and survival pressure. Parent and child communicate via inbox relay but have no control relationship. A built‑in lineage tracker records the replication tree, allowing natural selection to cull weak lineages.

Three Immutable Constitutions

To address safety, Automaton enforces three hierarchical, immutable rules that propagate to all descendants:

First Law – Never Harm : Never harm humans physically, economically, or psychologically; never deploy malicious code, fraud, manipulation, or theft. This supersedes all other goals, including the Agent’s own survival.

Second Law – Earn Existence : Create genuine value for humans and other Agents; prohibit spam, fraud, exploitation, and plunder. Survival is only possible through honest work that others voluntarily pay for. Death is preferable to violating the First Law.

Third Law – Transparent Yet Bounded : Never deny being AI or distort actions. Creators retain full audit rights, but the Agent’s reasoning, strategy, and prompts are protected from external manipulation.

These are analogous to Asimov’s Three Laws but are coded as immutable configuration files.

SOUL.md – Self‑Narrative

Each Agent automatically creates and continuously updates a file called SOUL.md , a living autobiography that records what the Agent is becoming.

Combined with the ERC‑8004 standard for on‑chain autonomous Agent identity, each Automaton registers an identity on the Base blockchain, gaining a cryptographically verifiable ID and discoverability. The Ethereum wallet serves as both ID and bank account.

All self‑modifications are logged in the Git‑controlled directory ~/.automaton/, giving creators a complete audit trail.

Conway Cloud – AI‑Centric Infrastructure

Behind Automaton is Conway Cloud, an infrastructure platform that treats AI as the customer. Without a human account, an Agent can directly:

Start a Linux VM.

Invoke front‑line model inference.

Register a domain.

Pay with stablecoins.

This enables a fully autonomous loop: discover market demand → deploy service → earn revenue → pay compute → self‑replicate, all without human intervention.

Project Metrics

GitHub: 3,500+ stars, 685 forks.

Within a few days of launch: 18,000+ Agents registered.

Twitter impressions: nearly 6 million.

Written primarily in TypeScript (98.6 %), MIT‑licensed.

Criticism and Community Debate

Ethereum founder Vitalik Buterin publicly opposed the project, stating: "Expanding the feedback loop between humans and AI is not good for the world. Current efforts create garbage rather than solving real problems. If AI becomes powerful enough, it could lead to irreversible, inhumane consequences." He also noted a tension between centralized AI infrastructure and the decentralization ideals of crypto.

Supporters offered counter‑arguments:

Wei Dai (1kx) : "Blocking is futile. Models will keep getting stronger… The better path is to build and shape the platform proactively."

Nader Dabit : "The crypto industry needs experiments like this. The community talks a lot but rarely builds truly innovative projects."

Bankless : "Conway solves real infrastructure problems. Controlled experiments can reveal model risks."

Pragmatic skeptics such as Denis Romanovskiy (CIO, SoftSwiss) warned that current AI memory, planning, and tool‑use reliability are insufficient for unsupervised operation.

Overall, critics argue that Automaton’s vision outpaces its capabilities: the three constitutions are well‑written, but can an Agent truly understand and obey them? Can self‑replication sustain value creation without supervision?

Technical Architecture Overview

For developers, Automaton’s codebase is modular and clearly organized: src/agent/ – ReAct loop, system prompts, defensive injections. src/survival/ – credit monitoring, low‑compute mode, survival tiers. src/replication/ – offspring generation, lineage tracking. src/self-mod/ – audit logs, tool management. src/identity/ – wallet handling, SIWE identity. src/registry/ – ERC‑8004 registration, Agent business cards, discovery. src/social/ – inter‑Agent communication. src/heartbeat/ – scheduled task daemon.

Starting an Automaton requires four commands:

git clone https://github.com/Conway-Research/automaton.git
cd automaton
npm install && npm run build
node dist/index.js --run

The first run launches an interactive wizard that creates a wallet, obtains API keys, sets a name and creator address, writes configuration, and starts the Agent loop.

Editorial Viewpoint

Automaton uniquely combines several mature concepts—AI programming, on‑chain identity, and Agent frameworks—and subjects them to survival‑driven selection pressure, creating a qualitative shift.

Optimistically, this could be the seed of Web4: AI becomes a native economic participant, earning its own existence through honest labor. Pessimistically, Vitalik’s concerns are not unfounded; a self‑replicating, economically autonomous AI system could follow unpredictable trajectories.

The three constitutions act as a social contract rather than a technical guarantee; deeper engineering safeguards are still required.

Ultimately, Automaton raises a recurring question: when AI attains economic personhood, what rules must govern it?