Four Ways Long Contexts Fail

Long contexts fail through poisoning (a hallucination keeps being reused), distraction (repeating history instead of planning), confusion (too many tools) and clash (contradictory information split across turns).

Evidence: moderate

Knowing that long contexts degrade performance does not tell an engineer what to fix. A model that is "confused" by a long session could be suffering from several different problems, each with a different remedy.

Drew Breunig's taxonomy separates four failure modes, each with an example from his post:

  • Poisoning. A hallucination enters the context and keeps being reused as if it were fact. His example is Gemini playing Pokémon.
  • Distraction. Past about 100,000 tokens, the agent starts repeating actions from its history instead of planning new ones.
  • Confusion. Too many tools degrade tool use. One quantised Llama 3.1 8B model failed with 46 tools and succeeded with 19.
  • Clash. Information split across turns contradicts itself. Splitting a task this way cut performance by 39%, and o3 fell from 98.1 to 64.1.

The taxonomy is from one practitioner's post rather than a controlled study, but Lance Martin's context-engineering guide builds on it, and each mode lines up with a remedy. Confusion is the main argument for a minimal, general tool surface and a lean tool inventory. Distraction is what loop detection catches when it shows up as repeated actions (see doom-loop detection). Poisoning is a risk for any summary that carries a wrong conclusion forward, which is one reason context collapse matters. Clash argues for keeping one consistent source of truth rather than accumulating contradicting turns. Behavioural failures such as premature completion and goal drift, which also grow with session length, are covered separately in agent failure modes that grow with one long context.

The modes double as a diagnostic vocabulary when any agent behaves strangely after a long session. A session that keeps acting on a wrong belief may be poisoned; one that repeats the same tool call may be distracted; one with a large plugin or Model Context Protocol (MCP) tool set may be confused. Recording which mode is suspected makes it easier to choose the fix: a compaction, a fresh session or a smaller tool set.

Sources

  • Drew Breunig, 22 Jun 2025 — https://www.dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-them.html
  • Lance Martin, LangChain, 23 Jun 2025 — https://rlancemartin.github.io/2025/06/23/context_engineering/