State of Context Engineering 2026: 6 failure modes and 6 fixes
dbreunig · x · 2026-11-03
Drew Breunig's AGNTCon talk "Context Fails & Fixes" serves as a state-of-context-engineering address, ahead of his O'Reilly book The Context Engineering Handbook.
Failure modes:
- Context Poisoning: hallucinations enter the context and get repeatedly referenced
- Context Confusion: superfluous info degrades response quality
- Context Distraction: overlong context makes the model ignore base knowledge
- Context Clash: new info/tools conflict with existing context
- Context Inertia: the model keeps pursuing a hypothesis it should abandon
- Fighting the Weights: your ask contradicts training tendencies
Tactics: Context Augmentation, Quarantine (isolate subtasks in dedicated threads), Summarization, Tool Loadout (load only relevant tool definitions), Pruning, and Offloading (store info outside the context).
Core thesis: the job is knowing when to be deterministic vs. probabilistic; models aren't people, and their construction creates unique failure modes. The talk cites research including Anthropic's Memento, Microsoft Research's Fork-and-Flush, Stanford's Stateless Language Agents, and UC Berkeley's Scaling Discovery Requires Forgetting.
More from coding & agent
- DeepSeek releases all 7 demo prompts for Harness Creator Mode plugins — rohanpaul_ai · 2026-11-03
- Ex-ByteDance engineer on Coze shutdown: workflows killed by models that plan — dotey · 2026-11-03
- Apple's iPrefBench tests personal agents on 130 underspecified iOS tasks: infer or ask? — hhsun1 · 2026-11-03
- Yale lab open-sources a Claude skill that makes publication-quality schematic figures — CatAstro_Piyush · 2026-11-03
- Prized launches Infinite Storage: boxes scale to 40TB at local-disk speed — thisguyknowsai · 2026-11-03
- Agent-era take: if AI reads your git history, the rebase-vs-merge debate stops mattering — davidcrawshaw · 2026-11-03