Rosso Cyber Cafe
An independent engineering log about AI agent reliability. The method is simple and unglamorous: take an agent system, break it on purpose in a way that could plausibly happen in production, and measure exactly what it did. Then write down the numbers — including the ones that make the agent look bad. Tools that come out of that work get shipped here, open source.
Latest notes
Did My Prompt Make the AI Agent Refund Twice?
I reran the same lost-response experiment 90 times with three retry policies. Under this configuration, the answer was yes.
I Made an AI Agent Refund a Customer Twice
A lost MCP response turned one $200 request into $400 and exposed why a tool-call trace is not a record of what actually happened.
Tools
once — a CLI that tests whether an AI agent can perform an irreversible side effect twice when a tool call’s result is lost. It drops the response to a committed write, watches what the agent does next, and reports whether the effect happened once or more than once.
once →Follow
New notes go out over RSS and byemail. No tracking, no popups, nothing to close.