Givi Pataridze

An agent cannot look at a screen, so a script becomes its eyes, and a broken script does not throw: it reports success. Five versions of my phone-width layout check, what broke each one, and the rule I now apply to any check: return a verdict and evidence that it observed.

The flat-rate subscription hit its weekly ceiling two days early, so I bought $250 in credits and turned the machine back on. Seven hours later they were gone, and I had 17 PRs, a real number for cost per PR, and a system that built its own cost controls.

189 pull requests opened, 163 merged, and hands-free merges rose to 31% while the reviewer got stricter, not looser. Throughput halved, which taught me the week's real lesson: an autonomous coder is bounded by its backlog, not its capacity.

In seven days Autopilot opened 97 pull requests and merged 87, median 39 minutes from open to merge. About one merge in five now happens with no human action at all. Nine escalations, two escaped defects, and a bug of the week that was, again, silent success at doing nothing.

Autopilot is an event-driven loop on GitHub webhooks and headless Claude Code sessions: a label starts an implementation, a failed CI run starts a fix, and a second AI reviewer gates every merge. Here is how it works, the guardrails, and the bugs that almost broke me.