The State of Context Engineering

219 points · 56 comments on HN · read original →

Points and comments are a snapshot, not live.

Long context windows introduce failure modes that require careful context engineering.

Drew Breunig's talk at AGNTCon outlines how long contexts in LLMs (up to 1M+ tokens) cause four failure modes: context poisoning, distraction, confusion, and clash. Examples include Gemini 2.5 Pro's Pokémon agent fixating on giving tea instead of lemonade, and research showing models follow contradictory examples in long histories. He advocates six tactics (augmentation, tool loadout, quarantine, pruning, offloading, summarization) and demonstrates simple DSPy harness code for tool loadout using a fast decision model.

What commenters are saying

Commenters largely confirm the degradation problem: several users report models getting "dumber" as context grows past 200-300k tokens, citing lost-in-middle effects and repeated behaviors. A camp argues modern models handle longer contexts better due to RLVR training, with one user noting no significant drop until 700-800k. Practical tips include auto-compaction before cache expiry, manual checkpointing via handoff files, and using provider-native compaction for reasoning traces. A few question whether decision-model pruning is worthwhile given cache invalidation costs.