The 80% Rule: Why Your AI Agents Should Only Speak When Confident
We've all been there: You ask your AI coding assistant for a solution to a tricky bug. It responds instantly, with absolute certainty, providing a code snippet that looks perfect. You copy it, run it, and... nothing. Or worse, a new error.
The AI wasn't lying to you. It was hallucinating. It was "confidently wrong."
In our Agentic Development Principles, we called this The Corollary of Confidence-Qualified Output. But in practice, we just call it The 80% Rule.
The Corollary of Confidence-Qualified Output was withdrawn from the Agentic Development Principles. A confidence score the model reports is its own estimate, not a verification (The Corollary of Confident Hallucination); an instruction to stay silent below 80% is a prompt that stronger signals outvote (The Principle of Interpretive Competition); and fact-checking less because the agent sounds sure is how vigilance decays. To have an agent surface its uncertainty, require it in the output's structure: The Corollary of Externalized Confusion.