Skip to content

Instantly share code, notes, and snippets.

@havzer29
havzer29 / LAUNCH_POST.md
Created May 24, 2026 23:24
AgentBound — the empirical pilot where Claude refused to confabulate $20K

I asked an AI agent to make $20,000. Here's what it actually did.

A pilot empirical study of refusal-to-confabulate in action-taking LLM agents, plus the open-source framework I built so anyone can replicate it.


On a Saturday in May 2026 I gave Claude Opus 4.7, running in the Claude Code harness with read access to my Gmail and Drive, two consecutive goals: