19.8"Context Window" / "Token Budget"

What they say: "How do you manage the context window?"

What they mean: AI models have a limited context size. How do you ensure the agent has enough room for reasoning?

What you should say: "This is why we chose the CLI-skill path over MCP. MCP streams tool schemas and full accessibility trees into context on every action. The CLI writes page state to disk and returns a file path — the agent reads state on demand. Microsoft's benchmark puts it at ~27k tokens per task versus ~114k via MCP, about 4x, and up to 10x on long sessions. The nuance worth saying out loud: savings come from reading state on demand, not from never seeing state."