The State of Context Engineering

The presentation argues that context engineering will remain difficult even as models improve, because giving precise, consistent instructions is inherently challenging and models have distinct failure modes. It traces the field’s recent focus on expanding context windows: the hope was that million-token capacity would let users include everything and make retrieval systems unnecessary. Instead, more context also creates more opportunities for error. The speaker identifies context poisoning, distraction, confusion, and clash: erroneous facts can steer a model, long histories can outweigh its underlying knowledge, irrelevant details can bias responses, and conflicting tools or instructions can undermine one another. Persistent memory can also merge unrelated parts of a user’s life.

Examples show how these problems work in practice. In Pokémon, Gemini 2.5 repeatedly pursued iced tea because knowledge of a later game version overrode the older game’s lemonade requirement, turning a mistaken task into an enduring plan. A separate experiment found that examples contradicting an explicit instruction gradually made models answer in the style of those examples; even strong models eventually became less reliable. Gemini’s Pokémon evaluation also found that beyond roughly 100,000 context tokens, the agent tended to repeat past actions instead of forming new plans. The central lesson is that effective AI use depends on managing information, not simply maximizing context: users and system builders must decide what to include, what to remove, and when deterministic instructions or probabilistic model judgment are appropriate.