Systems | Development | Analytics | API | Testing

I spent the day at NVIDIA thinking I’d be learning about all the newest models and RAM throughput achievements. Instead, a version of the future started to crystallize where AI really is good enough, pervasive and cheap(er). Let me share some of my observations and conclusions.

Every new signup posts a message to our Slack. The format is dull and reliable: Overnight this week one arrived like this: That trailing nothing was the entire incident. No error logs, no alerts. A returning user had signed up, our signup service had attached them to a tenant we deprovisioned back in December, and the only symptom in the whole company was a Slack message that ran out of words.

Ask an AI for K-pop concert advice without saying the group, city, date, or budget. It may confidently send you to a BLACKPINK tribute night in Cleveland with a $400 resale ticket. The AI was plenty confident. It just had nothing real to go on. That is exactly what happens when developers test AI applications with invented traffic. The test may look reasonable. The result may even pass.

The pitch for AI coding was speed. Claude Code, Copilot, Cursor, whatever you’re running, they all generate business logic faster than you can review it. That part is real. But look at what happens after the code gets written and the numbers get ugly. CircleCI’s 2026 State of Software Delivery Report found AI drove a 59% increase in average throughput.

We shipped v2 of a small products API on a Thursday. Green CI. Green replay. The new search endpoint worked. I went home feeling competent. Friday morning I ran the same traffic against both builds with proxymock and compared the SQL. v2 had added 80 queries on the same HTTP script. A per-product audit COUNT was firing inside the list handler. A startup migration had run ALTER TABLE and CREATE TABLE audit_log. Total DB time was up 70 ms on a demo that should have been boring.

A customer emailed on a Tuesday: checkout hung for ten seconds. I opened our tracing tool, punched in the time window, and got nothing. The trace was sampled out. We keep 1% of traces, like most shops with real traffic do. The one request that actually mattered was in the 99% we threw away. I spent twenty minutes admiring our observability stack before admitting it couldn’t answer a first-grader’s question: what happened to this person? Here’s what I know now.

It was 2am and I was paying for the privilege. Something was on fire in production, and I’d done the modern thing: I pointed an AI agent at it. It ingested the dashboards. It read the logs. It walked the traces. Then it handed me back a beautifully formatted paragraph that said, in effect, “latency is elevated on the checkout path.” I knew that. The page told me that.

In the first experiment, I wanted a baseline: if an AI coding agent gets the same production signal a human would get, can it fix bugs in a codebase it has never seen? Yes, but only when I gave it better context. With only an alert, the agent passed 51% of the runtime tests. When I added captured traffic, the actual request and response for the failing call, it climbed to 77%. This post is the second pass.