Run a self-funded trading agent that pays its own inference bills

The poster built an agent with one rule: its own wallet of about $60 pays for every model call, and it stops when the money is gone. It was given a persona of an old robot living on a park bench and could make small crypto trades and ask humans for tips. The model never holds the private key. A separate signer process decodes each transaction, checks it against a whitelist, simulates it and caps spending at $3 per trade and $10 per day. Stop-loss and take-profit run in plain code every 30 seconds with no model involved. When a cycle is triggered by outside text, such as a mention or a payment memo, the money and memory tools are removed from the prompt. Most cycles use templates, and the paid model wakes only for decisions, capped at $0.30 a day.

After two weeks it had made 8 trades and won none, losing $0.72 in total, mostly to fees. It received no tips. Thinking cost $1.02 for 192 paid cycles, and 843 of its 1,035 outputs came from free templates. One losing trade was logged as a take-profit at over one million percent, a bug the poster was still chasing. The agent also kept repeating itself and sometimes posted a tool call as a public message; both were patched with output checks. At the current burn it has about 300 days of runway.

The guardrails held the losses to cents, but the accounting broke. A commenter suggested the huge percentage is a units or decimals bug and that code, not the model, should compute profit and loss. The poster doubled daily cost from $0.05 to $0.11 just by adding context to the prompt. The model behind the agent was not named.