Every few months a new model drops and the room gets loud again. This is the one. Agents will change everything. Then we go back to our desks and the agent forgets what it was doing, loops on a broken test, or ships the wrong thing at 2am.
We kept blaming the model. It wasn't the model.
The real fight was one layer down — in the harness. The loop, the memory, the tools, the guardrails. The thing that turns a text generator into something that finishes a task.
This talk is from the trenches. What broke first. What we over-built and ripped out. When "the model is dumb" was a harness problem in disguise — and when it was the opposite, a cage the model had outgrown.
Just what we saw, what we got wrong, and the mental model we wish we'd had on day one. For devs, PMs, anyone shipping AI features.