Contract2Tool tackles a practical weakness in tool-augmented agents: tool schemas usually say how to call an API, but not enough about when the call is valid or what effects it should produce.
The paper proposes learning preconditions and effects for tools, giving agents a richer contract before they act. That could reduce invalid calls, bad assumptions, and brittle tool chains.
The idea fits a broader shift in agent engineering. As models get more capable, reliability increasingly depends on better interfaces, constraints, and execution checks around the model.