Developer-entrepreneur Theo detailed a 48-hour debugging saga on his podcast Theo - t3.gg, exposing limits in AI coding agents. His open-source terminal T3 Code, used by 120,000 people, was burning 50% GPU process CPU at full resolution on a 5K display despite being a text-heavy chat interface. Three AI models — Codex, Soul, and Fable — all failed to find the bug. One shipped a 10,000-line network rewrite that did nothing; another spent 30 minutes fighting Chrome DevTools before offering design advice Theo called "the worst guidance I've ever heard." The root cause, which Theo found himself by building a console toggle harness with agent assistance, was a single Tailwind pulse animation whose infinite opacity loop forced the GPU compositor to redraw at 120Hz indefinitely. The twist came after the fix: GPU usage stayed pegged because idle Claude.ai tabs were consuming roughly 10% of GPU each — the same failure class the agents could not diagnose. Late one night in July 2026, Theo noticed his $8,000 laptop was running hot. The browser's GPU process was chewing through 13–15% of the CPU at 720p, and spiking toward 50% at native 5K resolution. The episode is a calibration moment for anyone betting on autonomous coding agents. The models were fast, confident, and wrong. They were useful only when Theo stopped asking them to solve