My AI agent got dumber mid-session. I measured the context window before blaming MCP.
Scored daily by a customisable AI persona to surface the most relevant engineering leadership news.
Measures context window degradation in AI agents; directly addresses agent orchestration issues.
An AI coding agent's performance degradation mid-session was traced not to MCP tool definitions, which consumed only a small slice of the context window, but to accumulated conversation history that silently filled roughly a fifth of the window. The author measured token allocation by category before disconnecting MCP servers, discovering that idle connected servers cost little when the client defers schema loading. The fix is pragmatic: start fresh sessions after completing a task rather than carrying full transcripts forward, or have the agent summarize key state to preserve continuity without bloating context.