Agentic Coding in the Wild: Hype Meets Reality

A post titled 'Agentic coding notes from Galapagos Island' generated real engagement, and the comments were a mix of genuine excitement and skepticism. One commenter called it 'the beginnings of AI psychosis.' Another noted that Fable, an API-only coding agent, is expensive enough to run in a loop that you are essentially lighting money on fire. The megabyte-scale context window as a practical limit also came up as a frame for understanding where the hype has actually been right.

The through-line connecting this to the Leanstral 1.5 thread (a proof-generation model) and the Claude Mythos vulnerability spike thread is that AI systems are now being deployed in loops, for complex tasks, with real consequences. Bug finding, proof generation, and automated coding are all live experiments at this point.

The vulnerability thread noted that serious security bugs spiked around the Claude Mythos Preview release, with a commenter predicting a flood of disclosures once the responsible disclosure window closes. Whether those bugs are AI-introduced, AI-discovered, or both is the uncomfortable question.