Researchers at MIT documented 30 AI agents major labs are deploying. Only 4 had public docs saying what the agent does, what it can't do, and what happens if it breaks.