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
In an AGNTCon talk published October 24, Drew Breunig described how the push for longer AI context windows has exposed problems that larger windows do not solve. He argues that context engineering remains a challenge and involves deciding which tasks need deterministic controls and which can be left to probabilistic models.
Drew Breunig has published a talk on the state of context engineering, arguing that larger AI context windows have not removed the need to carefully manage what models are given. The October 24 report identifies context poisoning, distraction, confusion and clashes as failure modes that can affect model behavior, making the issue relevant to people building AI tools and agents.
Breunig’s talk, titled “Context Fails & Fixes,” revisits the industry push toward longer context windows. At the end of 2024, he says, a common hope was that models could be given enough context to hold vast amounts of information, reducing the need to select material or build more involved retrieval and agent systems. Model developers expanded windows from roughly 128,000 tokens to one million tokens, while some products and research efforts pursued even larger capacities.
Breunig argues that increasing the amount of information a model can receive also creates more opportunities for irrelevant, incorrect or conflicting material to affect its response. He describes context poisoning as information that should not be there, such as a mistaken fact, becoming difficult for the model to disregard. Context distraction occurs when a long context draws attention away from the model’s underlying knowledge. Context confusion involves unrelated details influencing an answer, while context clash describes tools, skills or other instructions in an AI system’s harness working against one another.
As an example of poisoning, Breunig cites the Gemini 2.5 technical report’s account of a model playing Pokémon. In the older Pokémon Red, a guard needs lemonade; in the later FireRed version, the item is iced tea. Breunig says the model repeatedly pursued tea while playing the older game, apparently drawing on knowledge of the later version. The example illustrates how information learned during training can interfere with a task even when the system’s immediate situation calls for something different.
Why Bigger Context Is Not Enough
The development matters because context-window size is not the same as reliable reasoning. A model may be able to accept more text, but that does not guarantee it will identify the relevant details, ignore incorrect ones or reconcile conflicting instructions. For organizations using AI agents, errors can come not only from a model’s answer but from the mix of documents, memory, tools and task instructions supplied to it.
Breunig’s account points toward a practical design problem: builders need to determine what information belongs in a model’s context and how to prevent unrelated material from shaping its behavior. His framing also distinguishes between deterministic controls, where a system follows explicit rules or checks, and probabilistic model behavior, where an answer is generated based on patterns. The talk’s title theme is that knowing when to use each approach is part of the engineering work.
The Pokémon example is a contained illustration, not evidence that all models will behave the same way in every application. But it shows why apparently minor conflicts between learned knowledge and current instructions can matter. In higher-stakes settings, a similar failure could affect a decision or action, depending on the system’s design and safeguards.
Products Worth Considering
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
LG UltraGear 27-Inch QHD OLED Gaming Monitor, 240Hz, 0.03ms, G-SYNC Compatible, FreeSync Premium Pro, HDMI 2.1
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
As an affiliate, we earn on qualifying purchases.
From Long Windows to Context Failures
Breunig traces the current debate to the period when long context windows were promoted as a way to place far more information directly into a model prompt. He recalls Gemini 1.5 Pro’s million-token context window as a prominent example of that shift, alongside discussion of a 10-million-token system. Llama 4 Scout later offered a 10-million-token window, he says, but inference challenges limited its appeal.
The expanded capacity changed what developers could provide, but it did not make selection and organization irrelevant. Breunig says the four failure modes he previously identified describe different ways context can undermine a task: wrong information can persist, excessive information can distract, unrelated facts can influence an answer, and instructions or tools can conflict. His talk presents these as recurring design problems rather than a claim that longer context has no value.
Breunig also connects the topic to the continuing difficulty of communicating instructions clearly. He says context engineering will remain challenging even as models improve, though the specific problems differ from communication between people because models are built and trained differently. The report is a talk transcript and commentary, not a controlled evaluation comparing systems.
“The job is knowing when to be deterministic and when to be probabilistic.”
— Drew Breunig, in his AGNTCon talk
How Broad Are These Failures?
The report does not provide a new benchmark or quantify how often the four failure modes occur across models. The Pokémon account is attributed to the Gemini 2.5 technical report, but Breunig’s article does not reproduce detailed test conditions or establish how frequently the behavior appeared. The scale of the risk in deployed systems therefore remains unclear.
The supplied transcript also ends as Breunig begins discussing research on how behavior changes as context grows. It does not include the demonstration’s results or the specific fixes promised in the talk’s title. The report consequently lays out the problem categories more fully than it documents remedies. It is also unclear how the failure modes vary by model, task, context length or system design.
Further Evidence and Practical Fixes
Breunig’s published transcript is a starting point for the next questions: which context-management methods reduce each failure mode, and how can teams test those methods against realistic tasks? His report signals that the talk covers fixes, but the available material here does not provide enough detail to assess particular interventions or their results.
For now, the account supports a measured takeaway for AI developers: expanding context capacity does not remove the need to manage instructions and information. More evidence would be needed to compare failure rates across systems and determine which safeguards work reliably. Breunig also invites readers to share feedback on the report and his forthcoming book, “The Context Engineering Handbook.”
Key Questions
What does context engineering mean in this report?
It refers to managing the instructions and information given to an AI model so that relevant material supports the task without introducing misleading or conflicting details.
What four context failure modes does Breunig describe?
He identifies poisoning (incorrect or unsuitable information), distraction (a long context drawing attention away from other knowledge), confusion (unrelated details influencing an answer), and clash (tools or instructions conflicting).
What happened in the Pokémon example?
Breunig says the Gemini 2.5 technical report described a model playing the older Pokémon Red but repeatedly seeking tea, an item associated with the later FireRed version. He presents this as an example of learned information interfering with the current task.
Does a larger context window make an AI system more reliable?
Not by itself. Breunig argues that larger windows allow more information to be supplied, but that information may still be irrelevant, inaccurate or conflicting. The report does not offer a quantified comparison of reliability by context size.
What remains unknown about the report’s findings?
The article does not quantify how often these failures occur, compare them across models, or provide enough detail to evaluate specific fixes. The transcript excerpt also ends before the research demonstration on longer contexts is described.
Source: dbreunig.com
