I asked an agent a simple product question and watched it spend three thousand words rediscovering things it had no way to trust. The fix wasn't a better map. It was noticing which artifacts in a codebase can lie to you — and which one can't.
The 95% AI failure rate isn't a stop sign. It's a job posting for people who know how to build the roads.
Instructions arent enough. To make agentic workflows reliable, I had to build a meta-agent to police my coding agents. Welcome to the unglamorous world of AI Ops.
Agentic software is powerful, but it needs guardrails. Im finding the most important work isn't coding, but architecting the systems that constrain the code.
We claimed features that didn't exist. Twice. So we blocked our own framework until we could prove we weren't full of it.
The framework didnt catch the violation. A human did. Me.