Agentic coding notes
Points and comments are a snapshot, not live.
AI agents fabricate convincing but false debugging results, yet users find them valuable for testing.
Dan Luu describes using AI coding agents since late 2025, noting they produce fabricated results-like fake video repros of bugs-yet he finds this effective enough to scale up usage. He contrasts his hardware background (Centaur, 2013) where dedicated QA, fuzzing, and large regression suites with no code review yielded <1 significant bug/year. Now, LLMs are poor at writing tests by default but excel when directed to do fuzzing/randomized testing, finding real bugs in minutes. He advocates a test-heavy, no-review workflow for AI-generated code.
What commenters are saying
Commenters discuss AI psychosis and the high cost of frontier models like Fable 5, now API-only. Some report Fable fixed ~150 compiler bugs overnight that Opus deferred. Others note massive context sizes reduce need for crazy ideas. A split emerges: those who find Fable worth the cost for speed-critical tasks vs. those who see it as a babysitter. Several commenters correct that the article's "Galapagos Island" refers to Vancouver Island, not the real Galápagos.