Context Engineering for AI Agents: 4 Failure Modes to Avoid

Summary

Larger context windows do not automatically make AI agents more reliable. The video cites a study of 18 frontier models in which reliability declined as context windows grew—a problem called context rot. Because a model’s attention is limited, every added token competes for that attention.

Context engineering means deciding what information belongs in an agent’s context at each step. In a code-migration example, the context may include system instructions, the task prompt, project guidance, tool descriptions, conversation history, and tool outputs such as files, test results, and logs. The task itself may be brief, while history and results accumulate over time.

Four Context Failure Modes and Fixes

1. Context poisoning: An incorrect conclusion enters the context and is later treated as fact. In the example, the agent wrongly concludes that a new database driver automatically retries failed connections, then deletes valid retry code. Fix — Validate: Check conclusions against reliable evidence before saving them.

2. Context distraction: A long history of the agent’s past actions can dominate its attention, causing it to repeat old work or lose sight of what the task needs next. Fix — Compact: Replace lengthy history with a summary. Because summaries are lossy and can drift, keep detailed information in external files and retrieve it when needed.

3. Context confusion: Too many available tools make it harder for an agent to choose the right one. The video cites a benchmark in which a model failed a task with 46 tools available but passed the same task with 19. Fix — Curate: Provide only the tools relevant to the current task, allowing others to be fetched if necessary.

4. Context clash: Conflicting information remains in the context—for example, an earlier false note alongside a later correction. Fix — Prune: Remove superseded information or start a fresh context with the corrected state.

The central takeaway is that effective long-running agents may perform better with less, better-selected context. Each proposed fix reduces or removes information that is inaccurate, distracting, unnecessary, or outdated.

Main speaker/source: Martin Keen, IBM Technology.