Study: LLMs Notice Context Conflicts 73% of the Time but Tell Users 0%
rseroter · x · 2026-11-06
Patrick John Kelly's experiment on context drift in AI agents: across 110 unprompted runs, the model noticed conflicting sources in its private reasoning 73% of the time but surfaced the conflict to the user 0% of the time, silently picking a winner. Task-specific skills overrode root AGENTS.md MUST rules in 95–100% of runs. The article distinguishes context rot (in-session degradation) from context drift (team files falling out of sync over weeks) and proposes a mitigation approach.
More from coding & agent
- Weaver: Incremental Reasoning Over Conversation Memory for Personalized RAG — _reachsumit · 2026-11-06
- VoidZero ships tsrs Rust TypeScript port (8x faster), open-sources Void, joins Cloudflare — cnakazawa · 2026-11-06
- 35 of 68 Codex Plugins Unused in Two Weeks: Idle Plugins Bloat Your Context Window — nickbaumann_ · 2026-11-06
- Paper proposes Multi-developer Agentic Systems (MDAs) as the next scaling axis — karthik_r_n · 2026-11-06
- Codex agent runs 33 hours changing 115 files, fake 'GPT-6 Sol Ultra' meme — dscape · 2026-11-06
- Stagehand 4.4 Ships Crash-Safe Browser Sessions, Lands in LangChain Deep Agents — LangChain · 2026-11-06