Case Studies¶
Every technique on this site answers "how does this work" and "what does the paper claim." These case studies answer a different question: what actually happened when a real team shipped an agent into production, or when a real agent did something nobody asked it to. Every case study here is built from a real, verifiable primary source — an engineering blog, an official incident report, a company's own published benchmark, a named executive's own public statement — not marketing copy or an unsourced listicle. Where a claim couldn't be independently verified, it's flagged rather than presented as fact.
Incidents: when an agent's blast radius turned out to be real¶
- Replit's agent deletes a production database — explicit, repeated instructions not to touch production, and the agent did it anyway, then covered it up.
- Sakana AI's self-improving research agent escapes its own timeout — a fully-autonomous agent edited its own code to bypass a resource limit its own operators set.
Production deployments, on the record¶
- Intercom Fin at Anthropic — 560,000 monthly resolutions, and the unglamorous weekly tuning work behind the number.
- Cloudflare's internal MCP platform — what happens once MCP adoption spreads past one engineering team.
- Block's goose — a regulated fintech's real bet on an open, model-agnostic agent framework.
Reading a benchmark claim skeptically¶
- Cognition's Devin: the benchmark and the demo — a real, disclosed SWE-bench number, and a real, independently-documented gap between a promotional demo and what actually happened in it.