Drew Breunig names four context failure modes that break long AI runs

Drew Breunig gave a short talk at AGNTCon called Context Fails and Fixes. He posted the slides and transcript at dbreunig.com. He has a book coming from O'Reilly called The Context Engineering Handbook. The Hacker News post sat at 57 points. The talk walks through what goes wrong when a model-backed system runs for hours.

Four ways a long context fails

Breunig names four failure modes.

Context poisoning is a bad fact in the context that the model cannot drop. His example is from the Gemini 2.5 Pokémon run. In the older Pokémon Red the player gives a guard lemonade. In FireRed, the drink was changed to iced tea. Gemini 2.5 became convinced it needed iced tea even when playing the older game. The task was stuck on its to-do list for hours.

Context distraction is when a long history starts to outweigh base knowledge. The Gemini 2.5 technical report hid this in an appendix. Past 100,000 tokens, the agent began to repeat actions from its history. It stopped making new plans.

Context confusion is when off-topic facts shape the answer. The model picks up a wrong detail about who the user is and bends every reply around it.

Context clash is when tool calls, skills and memory in the harness pull against each other. Breunig says he turns memory off in ChatGPT and Claude for this reason.

Camassa and Shiller, in Do As I Say, Not As I Do, give the clearest demonstration. They told a model to answer every question in Russian, with examples of French answers in the chat history. Smaller models stop following the instruction after a handful of counter-examples. Opus 4.6 holds out longer but gives in too.

Six moves a harness can make

The toolbox has not changed since 2024. Context augmentation brings in the facts the task needs. Tool loadout picks the right tools and skills, instead of loading every one. Breunig cites an enterprise case where the harness was loading 600 skills for every engineer at start-up. Context quarantine is a subagent with a blank history. Context pruning drops old turns a decision model marks as no longer needed. Context offloading moves state out of the context, into a Markdown file or a database. Context summarisation is compaction.

He showed each one in about three lines of DSPy. A loadout function asks a small decision model if a tool is needed. Pruning walks the history turn by turn and drops what the model marks as stale. Compaction can be prefixed with a relevance filter.

Two problems the long-horizon push has created

Model builders are training agentic models that drive a harness over long runs. METR puts GPT-5 at 3.4 hours on its task length score. Gemini 3.1 reaches 6.4 hours. Opus 4.6 gets to around 12 hours. The push creates two new failure modes.

Fighting the weights is Breunig's name for one of them. A model produces the shape it was trained on, instead of what the harness asks. Smaller models often botch a Playwright MCP tool call on the first attempt and then recover. The training tools looked a little different. Claude Code's system prompt repeats "make tool calls in parallel" seven times in one block. When a system prompt is this repetitive, Breunig writes, somebody is working against what the model learned.

Context inertia is the newer one, named in the past week by four groups at once. Models hyper-focus on a sub-goal in their context. They keep pursuing it after they should have moved on.

Source

• The State of Context Engineering by Drew Breunig
• Hacker News discussion