When the world becomes the test suite: agent intelligence for the physical world
When the World Is the Test Suite: Agentic Intelligence for the Physical World
Coding agents are writing, testing and fixing programs for robots, laboratories and industrial controllers, but the code comes without a test suite, and the paper treats the physical world as the test suite. The discussion centers on three questions: writing programs, testing them, and learning from the results, involving robot policies, reward and training world generation, simulators, formal checking and skill libraries. The authors point out that writing programs is increasingly cheap while trustworthy evidence is expensive, and that progress will depend on continuous integration with the physical world: verification that is cheap, reliable and maintained alongside experience.