Summary
Elvin Aghammadzada from DataRobot challenges a common assumption in agentic AI development: that larger context windows solve the tool overload problem. Drawing on the “context rot” paper, he explains that performance begins degrading after roughly 25% of a model’s context window is consumed — meaning an agent connected to 15 MCP servers can exhaust over 100,000 tokens in tool definitions alone before a single user message is sent.
The talk introduces “skills” as an architectural response to this problem. Rather than front-loading every tool definition, instruction set, and data schema into the context at startup, skills enable progressive disclosure — exposing only lightweight metadata at runtime so the model can selectively load capabilities as needed, similar to how a database index works. This reduces context bloat by roughly 10x compared to naive MCP setups.
Aghammadzada also frames skills as a new competitive moat for enterprise AI platforms. Where SaaS companies historically created lock-in through switching costs and integration friction, skills shift the dynamic toward what he calls a “fluency moat” — the agent becomes more capable and contextually aware the longer a user stays on a platform, compounding value through experience rather than friction. He notes that documentation traffic from coding agents has jumped from 10% to 50% year-over-year, signaling how deeply LLMs are already embedded in developer workflows and why structured skill surfaces matter for future enterprise AI.
📺 Source: AI Engineer · Published July 20, 2026
🏷️ Format: Deep Dive
