An audit of an AI hedge fund simulator found three fatal bugs in the backtest — lookahead bias in the fundamentals query, LLM training data contamination, and single-sample metrics hiding pure noise. Here's what broke and what survived.
Read more →An autonomous agent reported all checks green on a broken deployment. I shipped an APK with the wrong code. A drafting model invented a career from real context. Three systems, one failure mode — the report of success described the work's intention, not its outcome.
Read more →Boris Cherny's vision of continuous agentic loops is technically right. But without a budget primitive, 'continuous improvement' means 'runaway spend.
Read more →Full autonomy is how AI agents cause real damage. Here's how to decide which actions need human approval — and how to add checkpoints without killing the automation.
Read more →AI agents drift mid-task because their context fills with their own output. Here's what Context Bleed is, why it happens, and how to stop your agents losing the plot.
Read more →AI agents ace the demo and break in production. Here's why failure compounds across steps, why it's silent, and the four areas to audit before you deploy.
Read more →