TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Drew Breunig used an AGNTCon talk titled “Context Fails & Fixes” to review how AI context engineering has changed as models gained longer context windows. He outlined four failure modes—poisoning, distraction, confusion and clash—and argued that builders must decide which tasks need deterministic controls and which can be left to probabilistic models.
Drew Breunig used a talk at AGNTCon to lay out how AI context engineering has developed alongside longer model context windows, and why adding more information does not reliably solve problems with model behavior. His presentation, “Context Fails & Fixes,” identified four ways context can undermine a system: poisoning, distraction, confusion and clash.
Breunig’s account traces a shift in expectations from late 2024, when some industry discussion treated very large context windows as a possible replacement for retrieval-augmented generation, or RAG. The idea was that a model could receive all relevant material at once instead of relying on selective retrieval or more elaborate agent systems. Models moved from context windows of about 128,000 tokens to windows reaching roughly one million tokens, with Google’s Gemini 1.5 among the early examples he cited.
Breunig said those gains brought new room for errors rather than eliminating the need to choose and organize information. He defines context poisoning as incorrect or unsuitable information entering a model’s working context; distraction as a long context drawing attention away from other knowledge; confusion as irrelevant facts influencing an answer; and clash as tools, instructions or other parts of an AI harness conflicting with one another.
To illustrate poisoning, Breunig described a Pokémon benchmark example discussed in the Gemini 2.5 technical report. In the original Pokémon Red, a guard must be given lemonade; in the later FireRed version, the item is iced tea. Breunig said Gemini 2.5, while playing the older game, repeatedly pursued tea because information about the later game was present in its learned knowledge. In his account, the model then had difficulty abandoning that mistaken task.
Why More Context Can Backfire
The presentation matters to developers because context-window size is not the same as context quality. A system can accept more text and still perform poorly if that material contains conflicting instructions, irrelevant facts or incorrect assumptions. The challenge extends beyond prompt wording to the design of retrieval, memory, tool use and other components that decide what information reaches a model.
Breunig’s framing also highlights a practical division in AI system design: some tasks benefit from deterministic controls, while others depend on a model’s probabilistic output. His article’s central argument is that builders need to know when to rely on each. The presentation does not supply a universal method for making that choice, but it gives teams a vocabulary for identifying common sources of failure and investigating them.
For users, the examples explain why an AI assistant may confidently carry forward a wrong detail or blend separate projects when memory is enabled. For organizations deploying AI, such errors can affect reliability and make troubleshooting harder, particularly when several tools and instructions feed into one interaction.
Products Worth Considering
Amazon Kindle Scribe Colorsoft 64GB (1st Generation)
Apple iPad Air (4th Generation, 10.9-inch, Wi-Fi, 256GB) – Space Gray
Apple iPad Air 11-inch (M2), 512GB, Wi-Fi, Space Gray
As an affiliate, we earn on qualifying purchases.
From RAG Debate to Context Failures
Breunig presents his talk as a short account of the field’s recent evolution, rather than a claim that one technique has displaced another. In late 2024, he recalls, the “RAG is dead” argument centered on the prospect of putting far more information directly into a model’s context. The underlying hope was that sufficiently large windows would reduce the need to select information in advance.
He points to the move from approximately 128,000-token windows to million-token models as a major change in that period. Llama 4 Scout later advertised a 10-million-token context window, according to Breunig, but he says inference was challenging and the model did not gain broad traction. These examples are part of his talk’s historical account, not a comparative performance study.
Breunig also links the problem to how models are trained and optimized. They are not people, he argues, so users cannot assume they interpret instructions or combine memories in the same way a human colleague would. His forthcoming book, The Context Engineering Handbook, grew out of the concern that technical guidance can become outdated quickly; he told the publisher he remained optimistic that instruction clarity and consistency would stay relevant.
“The job is knowing when to be deterministic and when to be probabilistic.”
— Drew Breunig, in his AGNTCon talk
Limits of the Evidence Presented
The source is Breunig’s talk and accompanying lightly edited transcript, not a peer-reviewed evaluation of context engineering methods. The Pokémon example is attributed to the Gemini 2.5 technical report, but the supplied material does not provide the report’s full experimental setup, frequency of the behavior or a measure of how often it occurred.
Breunig introduces context distraction and says relevant research appeared after his earlier writing, but the supplied excerpt ends before he describes that demonstration or its results. It therefore does not establish how model behavior changes across context lengths, nor does it compare particular mitigation techniques. The talk’s account of model adoption and performance is also not supported here by independent usage data.
Further Findings and Practical Fixes
Breunig’s talk is titled “Context Fails & Fixes,” but the available source excerpt provides the four-part failure taxonomy and begins discussing distraction without including the rest of that section. Details of any specific fixes, evaluation results or recommendations are not available in the supplied material.
Breunig said he is preparing The Context Engineering Handbook, though the source does not give a publication date. The next useful evidence for readers and developers will be fuller descriptions of the proposed fixes and tests showing when they reduce failures without creating new conflicts. Until then, the talk’s supported takeaway is that larger context windows do not remove the need to manage what models receive.
Key Questions
What does context engineering mean in this report?
It refers to how builders select, organize and deliver instructions, information, tools and memory to an AI model so it can carry out a task.
What four context failures did Breunig identify?
He named context poisoning, where unsuitable information enters the context; distraction, where a long context draws attention away; confusion, where irrelevant facts influence output; and clash, where tools or instructions conflict.
Does a larger context window eliminate the need for RAG?
No such conclusion is established by the talk excerpt. Breunig describes the past expectation that large windows could make selective retrieval less necessary, then argues that long contexts bring their own failure modes.
What is the Pokémon example meant to show?
Breunig says Gemini 2.5 pursued iced tea in the older Pokémon Red game, where the guard’s item was lemonade, because the model had learned about the later FireRed version. He uses it to illustrate how prior information can lead a model toward a mistaken task.
What remains unknown about the talk’s findings?
The available source does not provide the full account of the distraction demonstration, quantitative results, or detailed evidence for proposed fixes. It also does not establish how frequently the Pokémon behavior occurred.
Source: hn
