Elvin Aghammadzada, DataRobot, mapped where a context window stops working: up to 40% full it's the smart zone, past that it's the dumb zone here's how to shrink an agent's context, from scratch: step 1 → measure your own edge, 2026 measurements put the start of degradation at 25% of the window, not near the end - a window is not a tank, it's a runway step 2 → count what's burned before the user says a word, an agent wired to 15 MCP servers spends over 100,000 tokens on tool definitions alone step 3 → load on demand, a skill's front matter is under 100 tokens and the body only arrives when it's needed, roughly 10x cheaper in context than plain MCP step 4 → put teachability on the buying checklist next to security and compliance, skills already ship natively in 26+ platforms step 5 → prune the junk, model-written skills perform worse than human-written ones and there is still no verification layer for them on August 1-2 OpenAI showed Astra: ten problems open for decades, cracked by coordinating multiple agents over long runs, with about $2,000 of tokens for all ten most people bolt on one more MCP server - he switches off everything the current step doesn't need bookmark & watch - this 26-min talk, then read the full context rot breakdown below ↓