Context rot is a real phenomenon where LLM agent performance degrades significantly as the context window fills up — accuracy can drop up to 30% going from 300 to 113k tokens. The post argues that most agent setups treat the context window carelessly, and offers concrete strategies: replace MCP servers with hand-crafted, minimal tools you control; use progressive disclosure in tool responses instead of dumping full documents or large JSON blobs; keep system prompts focused rather than exhaustive; and load only the tools a specific task requires, routing to specialized sub-agents instead of one bloated agent.
Table of contents
MCP: the biggest offenderTool responses are context tooYour instructions are context tooTool count is context tooEvery token should earn its placeSort: