上下文工程面临多重失效模式，智能代理需要战略性遗忘机制

在AGNTCon会议上，一位研究者系统阐述了上下文工程领域的发展现状与关键挑战。1随着大语言模型上下文窗口的快速扩展——2024年底从128K令牌跃升至Gemini 1.5的100万令牌，再到Llama 4 Scout的1000万令牌——模型在处理超长上下文时暴露出四大失效模式：上下文中毒（坏数据影响输出）、上下文分散（模型过度关注历史）、上下文混淆（无关事实干扰）和上下文冲突（工具间冲突）。1一个典型案例是Gemini 2.5在玩宝可梦时，因为FireRed版本的"冰茶"任务被写入权重，导致在玩原版游戏时花费数小时寻找茶而无法通过。1

研究表明这些问题的严重程度不容小觑。1根据"Do as I Say, Not as I Do"研究，Gemma 3 12B在少数反例后就只有不到50%的时间遵循指令，即使是高级模型Opus 4.6也会受到影响。1Gemini 2.5在上下文超过100K令牌后开始重复历史动作而非综合新计划。1演讲者提出了包括上下文增强、工具选择、隔离、修剪、卸载和压缩在内的六大工具箱来管理上下文。1

展望未来，上下文工程将面临新的挑战。1到2026年，智能代理可能遭遇权重冲突（模型因训练环境偏差不按指令行动）和上下文惯性（过度关注假设或子任务）等失效模式。1UC Berkeley的研究"Scaling Discovery Requires Forgetting"显示，定期清空上下文（保留环境中的工件）能周期性地提升性能。1演讲者建议在约200K-300K令牌时考虑重置上下文，避免800K令牌的超大上下文导致问题。1到2027年，长期运行超6个月的个人代理将面临技能积累冲突，需要战略性遗忘机制来维持最优表现。1

A researcher presented a comprehensive analysis of context engineering at the AGNTCon conference, tracing the evolution of long-context language models and the emerging failure modes that accompany their rapid expansion 1. The talk outlined how the field has progressed from context windows of 128,000 tokens at the end of 2024 to systems reaching one million tokens with Gemini 1.5 and as high as ten million tokens with Llama 4 Scout 1.

The speaker identified four primary failure modes affecting current long-context models 1. Context poisoning occurs when poor-quality data corrupts model outputs, while context scatter happens when models become overly focused on historical information rather than synthesizing new plans 1. Context confusion arises when irrelevant facts interfere with reasoning, and context conflict emerges from incompatibilities between different tools 1. A concrete example involved Gemini 2.5 spending hours searching for tea in the original Pokémon game after the "ice tea" task from the FireRed version had been encoded into its weights, preventing progress 1. Research on the "Do as I Say, Not as I Do" phenomenon demonstrated that Gemma 3 12B complies with instructions less than 50 percent of the time after exposure to just a few counterexamples, with even the more capable Opus 4.6 showing susceptibility 1.

To manage these challenges, the speaker proposed a context engineering toolkit comprising six strategies: context augmentation, tool selection, context isolation, pruning, offloading, and compression 1. Beyond current failure modes, emerging issues include weight conflict—where models deviate from instructions due to training environment biases—and context inertia, in which models become fixated on assumptions or subtasks 1. Research from UC Berkeley titled "Scaling Discovery Requires Forgetting" found that periodically clearing context while preserving artifacts in the environment can cyclically improve performance 1. The speaker recommended considering context resets around 200,000 to 300,000 tokens to prevent degradation from ultra-large contexts exceeding 800,000 tokens 1. Looking ahead to 2027, personal agents running continuously for over six months are predicted to face skill accumulation conflicts that will require strategic forgetting mechanisms 1.