Everyone assumes bigger context windows solve long-running AI agent problems. They don’t. A new design guide from AWS Samples reveals that agents break in predictable ways during extended tasks, suffering from context overflow and goal loss. The culprit isn’t the model’s intelligence; it’s the lack of a proper harness. When an agent runs for an hour with 200 tool calls, the sheer volume of data degrades performance regardless of window size. Chroma’s Context Rot report backs this up, showing that even top models like GPT-4.1 and Claude 4 become unreliable as input length grows. Attention mechanisms create quadratic relationships between tokens, meaning every added word depletes a finite budget. It’s not a bucket you can fill; it’s a resource with diminishing returns.

The real fix lies in context budgeting and offloading, not just summarization. LangChain’s Deep Agents and Anthropic’s Claude Code implement hard rules to keep the window clean. When a tool response exceeds 20,000 tokens, Deep Agents writes it to the filesystem and replaces it with a file path and a ten-line preview. Claude Code caps auto-memory at 25KB or 200 lines, deferring full tool schemas until they are explicitly needed. This approach prevents the original instructions from drifting into the middle of the context window, where recall is weakest. Manus reports that typical tasks require around 50 tool calls with a 100:1 input-to-output token ratio. Without offloading, that observation history clogs the system, causing the agent to forget its primary objective.

Subagents offer another layer of protection by isolating heavy lifting. Instead of one model trying to do everything, a coordinator spawns specialized workers. AWS AgentCore demonstrates this by running three browser subagents in parallel MicroVMs, sending only structured findings back to the analyst. This cuts runtime from potentially fifteen minutes down to four or six. Compaction handles what remains, but it’s risky. Lossy summaries often drop critical constraints. Claude Code mitigates this by re-reading recently modified files and reinjecting specific rules after compaction. Deep Agents goes further, structuring summaries to preserve session intent and next steps explicitly. If you’re building autonomous agents, stop chasing larger windows. Start building better harnesses.