I test AI agents for the failures that demos hide.
Over the past few months I've built Sentinel, a benchmark that puts MCP security guards through 6 classes of attacks in a sandbox, and I evaluate frontier coding agents on real bugs every week.
The biggest thing I've learned: an agent that looks great in a demo can still take unsafe actions, make up an answer, or quietly misread an ambiguous instruction once it meets real users.
If you're shipping an AI feature and want someone to break it before your users do, send me a message. I'll point out 3 issues I'd test first, free.
An AI agent fails halfway through a job. What should it remember when it tries again?
That question led me to build AgentTrace, a Python and FalkorDB project for keeping track of agent runs. It records task progress and decisions in a graph so you can inspect what happened and see where a run stopped. That’s especially useful when several agents share a workflow.
When you debug an agent run, what do you look for first: the failed step, the input that caused it, or the decision that sent it down that path?
The Forum: where the few who know are talking. Part of a short film series on how a company's knowledge turns into agents that actually work. Feedback from fellow builders welcome.