Vibe-coding debug note shows raw tool calls caught the real agent bug
andrey_kurenkov · x · 2026-07-28
A vibe-coding anecdote about an agent whose behavior was misinterpreted until raw tool calls showed the real cause: it was taking 2.3m strides instead of 1.4m, because the template only preserved the “where to stop” instruction and dropped the “how big a step to take” part.
The author says the simplest fix is to restore the missing step-size instruction in the taught template, so an offline probe can verify it before anyone puts the headset back on. The broader point: instrumentation is often correct while our interpretation is not.
More from coding & agent
- Supermemory pitches shared context transfer across Claude Code, Cursor, and Gemini — builditwithjoe · 2026-07-28
- NousResearch’s Hermes Desktop shows a borderless AI coding workspace — max_paperclips · 2026-07-28
- Teknium says Anthropic’s Opus 5 is worse than 4.6–8 in Hermes agent tests — Teknium · 2026-07-28
- A 2026 reading list flags agent harness overengineering, local Gemma 4 benchmarks, and AI infrastructure — rseroter · 2026-07-28
- Codex and Gemini end up arguing inside a client website build — DumpingSouptime · 2026-07-28
- youtube-skills adds YouTube transcripts and search APIs for AI agents — tom_doerr · 2026-07-28