Burak Karamık

It's measurement, not judgment. In my case platform is queryable, so the agent asks it the same question the app asks, then asks again with one thing changed — same record, different user, or the same data without the filter the screen applies. If it's already wrong before our code touches it, the answer is obvious. That's what stops a misdirected ticket, not the diagnosis itself so it is basically re-try loop. Unfortunetely it isn't bulletproof. I hit a case where the platform accepted a filter and quietly ignored it, so both queries matched and "not our code" looked proven. So a probe has to show it could actually see what it claims to check. And agreed on can't-reproduce — it's usually a data problem, which is why the agent has to try arranging data before it's allowed to say it.

Mostly the tool surface, not the reasoning. All those tools are what make a general agent good at general work; for me they were clutter, and every extra path was another way to do the same thing slightly wrong. Least subtle example: two browser providers installed, the wrong one quietly answering and recording nothing — runs looked fine and produced no evidence. No prompt catches that; a narrower tool surface does. Retrieval barely applies ( for now at least ), since I don't really much do it: a short always-loaded rulebook plus a pass over what earlier runs learned. And no clean before/after number — it's observational and objective :) . What survived narrowing was environmental, like verifying against a build without the fix. That took enforcement and evidence checks, not fewer tools.