Context Engineering Faces New Failure Modes

At AGNTCon this week, a researcher presented a short talk titled “Context Fails & Fixes,” which the source described as a kind of state of context engineering. According to the source at https://www.dbreunig.com/2026/10/24/the-state-of-context-engineering.html, the talk examined how context engineering emerged, where it has gone, and where it may go next. It matters because model makers are still targeting new capabilities, and those advances create fresh challenges for users and harness builders.

What happened at AGNTCon

The source reports that the speaker gave a short talk at AGNTCon and released slides plus a lightly edited transcript. The talk was titled “Context Fails & Fixes” and was framed as a sort of state of context engineering. The presentation moved through context fails and fixes, including how context engineering emerged and where it might go.

The immediate news value is not a single product launch. It is a field-level diagnosis. The source says the talk asked how context engineering emerged, where it has gone, and where it might go. As model makers target new capabilities, the source adds, they inadvertently create challenges that users and harness builders must overcome.

How context engineering emerged

The source traces context engineering partly to the end of 2024 and the so-called RAG is dead moment. The hope was that larger context windows could hold enough information for an AI system to work without retrieval. Model builders pushed long context capabilities. Gemini 1.5 Pro was promoted for its million-token context window and was said to have teased a ten-million-token model in the lab.

Llama 4 Scout later arrived with a ten-million-token context window, but inference was a challenge and it did not become widely used. The source notes that the field still received a million-token context model. The brief became: put everything in context and let the model sort it out.

That brief aged poorly. The source describes how long contexts are not a panacea and outlines failure modes such as context poisoning, distraction, confusion, and clash.

The failure modes getting attention

Context poisoning happens when unwanted information enters context, such as a hallucination or bad fact, and the model cannot move past it. The source gives the example of an agent playing Pokémon. In the original game, a thirsty guard needed lemonade; in FireRed, the change was iced tea. The Gemini 2.5 technical report described the model becoming convinced it needed tea and spending hours trying to solve the task.

Context distraction occurs when a growing context pulls attention away from the model’s base knowledge. The source references a paper called Do as I Say, Not as I Do. In it, models were given an instruction to answer in Russian, then shown contradictory examples. The more bad examples present, the less often the model followed the instruction.

Context confusion involves extraneous facts influencing a model in ways developers do not fully understand. Context clash involves tool calls, skills, and harness information conflicting.

Tools for fixing context

The source lists six tools that have become central: context augmentation via RAG and skills, tool loadout, context quarantine, context pruning, context offloading, and context summarization or compaction. These have moved into the harness.

The source uses DSPy as an example. An agent can be a while loop with a ReAct module. A small decision model can select tools. A loadout function can filter tools. A subagent can provide quarantine. Pruning can cut history. Compaction can summarize. The source notes prompt caching should be minded.

What this means for AI builders and Grafy users

Context engineering will remain hard, but in new ways. The source says models are not people, and their built-in traits create unique failure modes. For developers, context is now part of system design. In Grafy, the Code app (/code) is the most direct surface for testing how coding agents handle instructions, tools, and history, and the Graph app (/graph) can help trace how prompts, outputs, and model history branch. The source does not claim Grafy has solved these problems; it offers a toolkit that builders must apply.

The practical takeaway is to decide where determinism helps and where probabilistic generation helps. Context fixes should be evaluated, not assumed.

To explore this yourself, open Grafy Code: an AI coding studio.

Frequently asked questions

Is context engineering the same as prompt engineering?

Not exactly. Prompt engineering is a part of it, but context engineering covers broader context handling, including retrieval, tools, memory, pruning and summarization. According to the source, failures like context poisoning and distraction show why managing context matters. The source argues that as models improve, the challenge shifts from simple instructions to consistent, clean context. The answer is not only better prompts.

Why did long-context optimism fade?

After late 2024, the industry hoped larger context windows would make retrieval less necessary. The source notes the “RAG is dead” moment, where putting the whole internet into context was seen as a solution. Gemini 1.5 Pro emphasized a million-token window and teased larger models. Llama 4 Scout later offered ten million tokens, but inference limits hurt it. Experiments later showed long contexts could still fail.

How can teams use context fixes?

Teams can use context augmentation, tool loadout, quarantine, pruning, offloading and summarization. The source says these tools moved into harnesses. A demo with DSPy showed an agent loop, ReAct, a small decision model, tool loadout, subagents, pruning and compaction. Teams should filter skills, isolate agents, prune history and summarize context. The goal is to keep relevant information while avoiding noise and errors.

---

Source: Hacker News — front page — The State of Context Engineering