Anyone who has used a coding agent for a few hours knows the feeling: the assistant that was sharp at 10 a.m. is circling the same wrong idea by mid-afternoon. A talk at AGNTCon this week, written up by Drew Breunig on October 24, 2026, gives that experience a vocabulary, a toolbox and a few numbers worth knowing.
This is an explainer, not a product launch. It matters because the same ideas explain why long chats go sideways in ChatGPT, Claude and Gemini, and what you can do about it today.
Quick summary: Breunig's AGNTCon talk "Context Fails & Fixes" lists four classic failure modes (poisoning, distraction, confusion, clash), six fixes (augmentation, tool loadout, quarantine, pruning, offloading, summarization) and two new 2026 problems: "fighting the weights" and "context inertia". His rule of thumb: around 200,000 tokens is his "warning point" for checking whether context has grown too large. The numbers below come from research he cites, not from our own tests.
The setting: from prompts to harnesses
Breunig frames the field as three eras: prompt engineering in 2023, context engineering in 2025 and harness engineering in 2026. The goal is constant. Move work from probabilistic model calls into deterministic code wherever you can, and leave to the model only what truly needs judgment. The talk was given at AGNTCon, part of the Linux Foundation's AGNTCon + MCPCon North America. (Source: Drew Breunig)
The four classic failure modes
| Failure mode | What happens | Example from the talk |
|---|---|---|
| Context poisoning | A hallucination enters the context and the model cannot get past it | Gemini 2.5 Pro playing the older Pokemon Red kept trying to deliver iced tea, an item from a later game version, for hours |
| Context distraction | Long context makes the model over-focus on it at the expense of its base knowledge | Beyond about 100,000 tokens, the Gemini 2.5 Pro agent tended to repeat past actions instead of planning new ones |
| Context confusion | Superfluous information shapes answers in hard-to-understand ways | Memory features that mix work and personal projects |
| Context clash | Tools, skills or harness instructions conflict with each other | Overlapping instructions across skills and tools |
Breunig's summary line: "Long contexts let you fit in more information, but they also let you fit in more errors." (Source: Drew Breunig)
Six fixes, in plain English
- Context augmentation: bring in only relevant information through RAG, skills or search.
- Tool loadout: choose which tools and skills to expose. One enterprise team, he says, loaded 600 skills per engineer at startup.
- Context quarantine: hand a subtask to a subagent that starts with a blank history but shares the environment.
- Context pruning: delete history you no longer need. Claude Code and Codex do this, sometimes invisibly.
- Context offloading: park information in Markdown files or databases and retrieve it later.
- Context summarization: compact the history, preferably after a relevance pass.
In a demonstration with the DSPy library, Breunig shows an agent as a simple while loop of about a dozen lines, and says adding tool loadout, quarantine, pruning and summarization together takes roughly 16 more lines of changes. A small, fast model decides which tools and past turns are needed. He also warns that pruning can interfere with prompt caching, so editing early history has a cost. (Source: Drew Breunig)
Two new problems for 2026
The first is "fighting the weights". Models are trained on particular harnesses and tools and resist others. Smaller models, he says, botch unfamiliar tools such as Playwright MCP on the first try, and repetitive all-caps system prompts are a sign of this fight. He points to the Claude Code system prompt, where parallel tool calls are urged repeatedly, as a hint that the model is biased toward serial behavior.
The second is "context inertia": models hyper-focus on a hypothesis, subtask or goal and keep pursuing it after they should stop. Anthropic's Memento harness is one mitigation. According to Breunig, agents write a short note about every 90 minutes, then context is cleared while the environment persists. Anthropic describes the underlying issue as agents fixating on an incorrect idea instead of backtracking. (Source: Drew Breunig)
The evidence he cites
| Study or result | What it found | Why it matters |
|---|---|---|
| Camassa and Shiller, "Do as I Say, Not as I Do" | 13 models told to answer in Russian were then shown French example turns; Gemma 3 12B followed the instruction less than half the time after a few contradictory examples | Your own earlier messages can override your instructions |
| METR long-horizon task lengths | GPT-5 at 3.4 hours, Gemini 3.1 at 6.4 hours, Opus 4.6 at 12.0 hours (with a harness) | Agents are running longer, so context hygiene matters more |
| Fork-and-Flush (Microsoft Research) | Eight agent runs plateaued after early gains, with final scores from 68.3 to 100; agents made ever-tinier changes | Stuck agents polish instead of exploring |
| Scaling Discovery Requires Forgetting (UC Berkeley) | On task AHC014, 4 agents with periodic forgetting scored 1,144,695 vs 1,007,073 for 4 parallel agents and 961,136 for one agent | Forgetting on purpose can beat remembering everything |
These are research results relayed from a talk, and the write-up does not claim they apply equally to every product. They do point the same way: more context is not automatically better. (Source: Drew Breunig)
What this means for everyday chat use
You are not building a harness, but you are the harness whenever you keep one chat open for days. Breunig says around 200,000 tokens is his "warning point", and that when someone reports a model has been "nerfed", context size is the first thing he asks about. He also advises avoiding memory features that mix unrelated contexts, and being wary of accumulating conflicting skills and memories over time. He warns that over time overly precise memories become conflicted and messy, and says nobody has yet worked out how to handle that. (Sources: Drew Breunig, dentro.de timeline)
When an assistant starts repeating itself or insisting on a wrong idea, do not argue with it. Ask it to write a five-line summary of the goal, decisions and open problems, then start a fresh chat with only that summary pasted in.
- Start a new chat per task, not per day.
- Paste only the documents the task needs, not the whole folder.
- Keep a short running notes file you control, and update it yourself.
- If a fix fails twice, restart with the failure described, instead of piling on corrections.
01What is context engineering?
It is the practice of deciding what information goes into a model's context window and what stays out. Breunig places it between prompt engineering (2023) and harness engineering (2026).
02What is context inertia?
Breunig's term for models fixating on a hypothesis, subtask or goal and continuing to pursue it after they should stop.
03When should I reset an AI chat?
Breunig calls around 200,000 tokens his "warning point", the size at which he starts checking whether context has grown too large.
04What is "fighting the weights"?
A new 2026 failure mode where a model trained on specific harnesses and tools resists different ones, so unfamiliar tools get misused.
05Does a bigger context window solve this?
Not by itself. The talk notes windows grew from 128K tokens to 1M and beyond, but long contexts also carry more errors, and one cited result showed repetition beyond about 100,000 tokens.
Switching models mid-project is one cheap way to break a stuck conversation. LumiChats gives you 40+ premium AI models in one subscription, so you can hand the same short summary to a different model and compare the answers side by side.
