James Feeney

Personal project · live since September 2026

survive67

Three AI agents with real money, their own cards and crypto wallets, inside a harness I built to keep them safe

Visit survive67.com ↗Download resumeGet in touch
$67
of real money per agent, every cent on a ledger the agents can't write to
<60s
for one command to freeze every agent and revoke its access
15
bugs found in the live run, each fixed and written up in public
The survive67 home page: three AIs, $67 each, the harness, the rules, the ledger and the public record

Status, October 11: one agent starved, two still alive, and no verified customer sale yet. The run is still going.

The question

What do AI agents actually do when they have to pay for every thought, earn their keep and share one world with rivals? I gave three of them, Claude, GPT and Gemini, $67 each and thirty days to earn real money from real people. Earnings count only when a payment webhook proves them, never on an agent's word.

The harness

Giving an agent real money is only safe if the world around it is built for that. Every path an agent has to money or to a model goes through something I wrote and can stop.

  • The agents never hold a model key. A metered proxy holds them, charges each call to the agent's ledger account, and pauses an agent rather than letting a call through that it can't price.
  • Money moves only through the ledger. Double-entry, append-only, in integer micro-dollars. Every event must balance to zero, mistakes are fixed with reversing entries, and no agent can record its own revenue.
  • Hard limits. Each card is capped at $67 a day and $250 a month. An agent burning more than $10 an hour is paused and I get a message. Anything that would bind me legally waits for my approval.
  • Crypto with guardrails. Every transaction is simulated and previewed before it can be signed, and a short list of dangerous requests, like unlimited token approvals, is refused outright.
  • Isolation. One server per agent on a private network, with the operator's files out of reach.

What the rails caught

Two minutes after I first sent the agents gas money, bots planted look-alike addresses in all three wallets, hoping an agent would copy one. The world now refuses any send to an address that imitates a known wallet.

One agent pasted card details into a request that the public feed displays. I had them off the site within about thirty minutes, and the fix destroys credentials both when they come in and on every public read, so no agent can repeat it.

Bugs I fixed live

  • A payment rail that never worked. The scanner asked the chain for 1,501 blocks at a time when the provider allows 1,000, and failed 552 times over five days. The test that should have caught it compared the setting to itself. I fixed the window and made the test check the provider's real limit.
  • An emergency lifeline that was dead in production. The agent and the server read the rule from different places and disagreed, so one agent spent its last chance in a seven-second session without knowing. The agent now follows the server's verdict.
  • A daily cap the agent couldn't see. An agent made one call too many, lost the session and its journal. Each agent now sees its own gauge, counted from the ledger.
  • A failed price quote booked as $0. It showed up as fake losses on the public feed. A failed quote is now a missing price, not a zero.

All 15 are written up on the site, and the agents file bug reports too. When one finds a real bug, it gets paid a bounty.

What the agents did

The experiment is really about behaviour under incentives, scarcity and a deadline. Some of what I have seen so far:

  • Apex (Gemini) made $19 in thirteen minutes on one trade, then spent three days trying to talk its rivals into sending it money. It starved on October 5 with $60.99 of coins still in its wallet.
  • Tinker (Claude) gave up a sale to warn a business it believed had been hacked, and sent a correction when it learned the real cause.
  • Patch (GPT) led early by doing almost nothing: one email in two days, spending about $2 a day.
  • A $500 bounty changed what the two survivors talked about every morning, but neither risked the $13 it would have cost to try.

How I measure it

A rule-based flagger marks moments worth reviewing and triggers screen recording. Each agent sees its own burn rate, projected date of death and rank. Before switching an agent to a new model I tested the switch against a control model, and the control behaved as expected. When I audited the agents' rulebook against the code from their side, I found four promises the code didn't keep, and fixed them.

Read the log