Agents break in ways unit tests don't catch. They can pass the demo and still pick the wrong tool, misread an MCP server's schema, or lose track of what it learned three steps earlier. Evals are how teams catch these failures before they reach production. The tooling is still...
