Context Inertia: when LLM agents latch onto a wrong hypothesis and won't let go
rseroter · x · 2026-10-17
dbreunig names a new agent failure mode — Context Inertia: the model commits to a hypothesis, subtask, or goal in its context and keeps pursuing it after the evidence says to stop or change course.
Key points:
- It's part context poisoning (an error gets repeatedly referenced) and part context distraction (over-focusing on a bloated context), amplified by frontier models trained to relentlessly follow instructions.
- Anthropic's case study: its Memento harness addresses this — each agent works until a fixed timeout (1.5 hours in one experiment), writes a short note, and a fresh agent inherits the codebase, environment, and the note only. Example failure: a Qwen agent spent hours debugging package versioning when the real bug was an endpoint incompatible with the grader API.
- Two papers the same week (Fork-and-Flush and others) converge on the same fix: forgetting. Periodically flushing context and handing off via notes beats letting one agent grind on.
rseroter adds: this is exactly why it's smart to watch agents work so you can step in and redirect.
More from coding & agent
- Engineer's long-form guide: building your own personal software factory — maddiehfaulkner · 2026-10-17
- 45-year game coding veteran: agentic coding is the biggest productivity jump I've ever seen — IgorCarron · 2026-10-17
- PKU & Samsung: slow model coaching lifts fast agent's live game score from 13.51 to 24.19 — rohanpaul_ai · 2026-10-17
- Scale AI's AgentEnv RL environments framework lands on Hugging Face Hub — ben_burtenshaw · 2026-10-17
- CoPE adds one forecasting loss to GRPO for 2x sample efficiency in agent RL — youjiaxuan · 2026-10-17
- LeanGraphSearch mines Mathlib dependency graphs, lifting proof success to 81% — kaiwei_chang · 2026-10-17