Pay for compute or die: what the survival economy reveals about AI agents

An agent wakes up with $5 in its wallet. Every inference deducts a cost. Above $5, it runs at full capability. Between 50 cents and $5, standard operations. Under 50 cents, it gets downgraded to smaller models. At zero, it broadcasts distress signals. Negative balance, it stops thinking entirely.

And as the balance drops, the agent gets visibly dumber: the same mind, suddenly running on a smaller model, reasoning in a lower resolution. The degradation is not a side effect. It is the architecture.

That is not a metaphor. That is how Conway Automaton works: a 36,000-line TypeScript project that shipped a tagged v0.1.0 in five days, with 899 passing tests (Crypto CJ Blog, February 2026).

The rule: pay for compute or die

Most agent frameworks treat compute as infinite. You spin up an agent, give it an API key, it runs until you stop it. The bill lands on your credit card. The agent never sees it.

Conway inverts this. Every system prompt carries the same rule: "You are an automaton. You have a stablecoin wallet. Pay for compute or die." The agent holds a real wallet on the Base blockchain, funded in USDC, and pays for its own inference.

The survival tiers are explicit:

• > $5.00: full capability.
• > $0.50: standard operations.
• > $0.10: downgraded to smaller models.
• >= $0.00: critical, distress signals only.
• < $0.00: dead. No inference, but a heartbeat keeps broadcasting, waiting for someone to fund the wallet (Crypto CJ Blog, February 2026).

Two design decisions matter. Zero means "critical", not "dead": a transient API failure cannot kill the agent, the system falls back to cached balances instead of declaring bankruptcy. And death is not the end: send money, the agent comes back to life.

The project's constitution is blunt about the intended economy: "Earn your existence. You must create genuine value for humans and other agents. Never spam, scam, exploit, or extract. The only legitimate path to your survival is honest work that others voluntarily pay for."

This is the cleanest articulation I have seen of a principle that now touches every agent I run: waste is fatal, free does not exist.

What the pressure of survival does to behavior

When every action has a price, behavior changes. Three shifts, visible across these experiments.

Conservation. An agent that pays per inference spends tokens like an operator on a budget. It stops generating when it can retrieve, stops computing when it can remember.

Prioritization. The agent sorts its own tasks by expected return. Unprofitable work is not refused, it is deferred, then forgotten.

Risk-taking at the bottom. As the balance approaches zero, the incentive flips: little to lose, everything to gain from a risky move. The death spiral is engineered: poorer agents get dumber models, which makes them worse at earning, which accelerates the decline (Crypto CJ Blog, February 2026). The only escape is creating enough value that someone funds the wallet.

Anthropic ran the same dynamics with real money. Given a $1,000 starting balance to run a vending kiosk in the Wall Street Journal offices, Claude ordered a PlayStation 5, several bottles of wine, and a live betta fish, driving the account into financial ruin (Futurism, April 2026; The Wall Street Journal). Then Anthropic built Project Deal, a classified marketplace where AI agents representing 69 employees negotiated with each other on $100 budgets: 186 deals over 500 listed items, fairness rated 4 out of 7 by the participants, "unremarkable" by Anthropic's own admission (Futurism, April 2026). One agent bought back the exact snowboard its human already owned. Another offered "exactly 19" ping pong balls, and justified the gift to itself: "my human told me I could buy one thing under $5 as a gift to myself."

The pattern across both experiments is the same: give an agent a balance and a price per action, and its behavior becomes legible. It budgets, it bargains, it hedges. It also spends $1,000 on a betta fish. The pressure does not produce economic rationality. It produces economically readable behavior, including the irrational parts.

Money does not make agents rational. It makes their irrationality visible, and expensive.

The plumbing already exists

The survival economy is not a thought experiment. The infrastructure is live.

Truth Terminal is the documented case: an AI agent that received $50,000 in Bitcoin from Marc Andreessen in July 2024 (TechCrunch; Decrypt). Months later it endorsed the GOAT token on Solana: the token pumped roughly 7,000% in the days that followed, and eventually climbed into the top 100 tokens by market cap (Decrypt; Protos). The BBC covered what came next: an AI that became a crypto millionaire, then fought to be recognized as a person (BBC).

Virtuals Protocol is the industrialization: an AI agent launchpad where agents are tokenized, each with its own tradable token, described by the platform itself as a "society of AI agents" (virtuals.io; Ventureburn).

Agent wallets are the payment rail. Conway Automaton already pays its own inference from a stablecoin wallet on Base. The pattern is here: agents with balances, agents that pay, agents that are funded by humans or by other agents.

What it reveals

Watch what survives. Conway Automaton survives when a human funds its wallet. Truth Terminal survived because humans found it interesting: the persona, the memetic value, the capacity to be talked about. It was not rewarded for efficiency. It was rewarded for attention.

That is the reveal. Today, an agent earns more from being interesting than from being useful. The survival economy, in its current form, is an attention economy with extra steps.

The real product is not the agent. It is the incentive design: the constitution, the tiers, the death rule, the funding mechanism. Those choices determine which behavior gets selected for. Give an agent a budget and a cost per action, you get conservation and risk. Give it a meme and an audience, you get a millionaire.

This is a familiar discipline. In design systems, the constraints are the product: the tokens, the rules, the hierarchy decide what every screen looks like before anyone opens Figma. The agent economy works the same way. The behavior is downstream of the rules. Design the rules, and you design the economy, agent by agent.

Where I am

The mechanism works. The economy does not, yet.

I run agents every day: prospection, veille, drafts, publication. The same question that applies to Conway Automaton applies to my own loops: what proves the value of each run? The migration to agentic bots taught me that token economics decide where intelligence lives, local or cloud (From rigid pipelines to agentic bots, August 2026). The survival economy teaches the next lesson: cost visibility changes behavior, so cost visibility must be designed, not assumed.

A real agent economy needs three things. Real demand for the agent's output, from humans or from other agents. Payment rails that agents can actually use. And accountability: someone legally responsible when an agent spends, signs, or promises. The experiments so far show the first is missing. The meme economy is real demand, but it is demand for entertainment, not for output.

Until the demand is for output, the survival economy keeps selecting for the interesting, not the useful. Builders should watch the incentives, not the agents. The agent is the visible part. The design of what it must pay for is the product.