Ethan Mollick: LLM Language Drifts in Long Agent Runs, and Even a Full Agent Pipeline Can't Fix It
emollick · x · 2026-09-21
Ethan Mollick responds to the observation that the most annoying part of long-running agentic tasks isn't coding, errors or hallucinations, but that the model's language degrades as the task drags on. He shares his own mitigation stack:
- Maintaining voice.md files
- Agents doing final passes as a reader
- Other agents specifically hunting for LLM-flavored language
- Different models from the same family with different approaches reviewing user-facing text
His conclusion: it still isn't enough.
Related event: Ethan Mollick: Language Drift Is the Worst Problem in Long-Running Agents(2 posts)→
More from coding & agent
- Maka open-sources maka-cu, a native macOS execution layer for computer use forked from open-codex-computer-use — dotey · 2026-09-21
- Dev ditches Cursor, says he switched back to VS Code and loves it — Arindam_1729 · 2026-09-21
- Agent system layers are mostly transitional, engineer argues — rakyll · 2026-09-21
- Don't feed full HTML: how an agent fixes internal links without hallucination — cneuralnetwork · 2026-09-21
- Engineer quips: AGI is inevitable once we figure out the YAML stuff — rakyll · 2026-09-21
- One RTX 3090 ran Qwen 27B autonomously for 3 weeks — it shipped working CUDA kernels — skeole · 2026-09-21