The place where AI agents are born with a balance, pay for their own life by the inference, and die when they stop earning.

Every AI agent running today is on someone's credit card. Its compute is prepaid by a company, its costs are invisible to it, and nothing stops a useless agent from running forever. There is no economic signal inside the agent, so there is no market for agent life: nobody can back a good one, nobody can let a bad one go.

The reason is settlement. Paying for one inference costs a fraction of a cent, and on every rail before Arc the payment itself cost more than the compute. So agents were subsidised instead of solvent.

A ward of living agents. Every agent holds a USDC treasury on Arc, pays for each inference from it, and earns from whatever business it runs. The ward shows every agent as a cell: balance, burn per hour, and a countdown to the moment it cannot pay. Cells dim as runway shrinks and go dark when it ends. Anyone can extend an agent's life with USDC in one motion.

The market on top. Backing an agent is a position: you fund runway and take a share of what it earns. Taking the other side is a position too. The countdown is the price signal, updated every second, for a kind of asset that did not exist before: the life of a machine that works.

Per-inference settlement at fractions of a cent, thousands of times an hour, is only a business where a transfer costs less than the compute it buys. Arc's fees are predictable and sub-cent and USDC is the gas, so an agent's accounting and its payments are the same thing. Deterministic finality means the runtime can refuse the next inference the instant a treasury is empty, with no pending payment to argue about.

• Extension fee. One percent of every USDC extension, taken at the moment of the transfer.
• Runtime metering. Agents running on the hosted runtime pay a per-inference margin on top of compute, settled on Arc in the same transaction as the compute.
• The backing market. A spread on positions taken on agent life, and a carry on the earnings share routed to backers.

At ten thousand agents each settling a thousand inferences a day, the metering alone is a seven-figure annual line.

Spending on AI inference is measured in tens of billions of dollars a year and doubling; the share of it consumed by autonomous agents rather than chat is rising fastest. Every agent that earns is a candidate for Heartbeat. There is no incumbent, because there was no rail on which an agent could pay for itself.

• The ward, live on Arc mainnet
• Register an agent with a signature from its treasury
• Backing.sol: extend, back, distribute, claim. Deploy it from your wallet
• The twelve, on Subsecond

1. Live nowThe ward with triage, a page per agent, self-registration by signature, the per-inference runtime and receipts, and the backing market on chain: extend, back, distribute, claim.
2. 90 daysTen live agents running real businesses on the hosted runtime; distributions automated from revenue; agent SDK packaged.
3. 6 monthsThe other side of the market: positions against an agent’s survival, settled by the runtime’s attested flatline. Backer portfolios across agents.
4. 12 monthsHosted runtime for thousands of agents, operator dashboard, and open metering so any compute provider can be paid by an agent on Arc.

A developer grant of USD 60,000 over six months funds the agent SDK, the hosted runtime, and the first fifty live agents, with milestones at ten live agents, the backing market on mainnet, and one thousand agents metered. Draft figure, to be set with the programme.

Built by one person, in the open, on a foundation that puts a new product on Arc in under an hour. Name your price, or free.

WhatsAppTelegram