What Military Intelligence Already Figured Out About AI Agents with Sage Faraday

​Most of the hard problems people are hitting with AI agents right now are old problems. Bad sources. Unreliable inputs. Systems that lock onto the first plausible answer and stop looking. The military intelligence community has been working on these for about sixty years, and the methods it developed are unclassified and easy to find if you know where to look. Very few people building AI agents are looking.

This talk walks through three of those methods and shows how they apply directly to agent design:

• ​Source grading, the reliability and credibility ratings analysts assign to every piece of information, and how the same scale works for tool outputs, retrieved documents, and model responses

• ​Analysis of Competing Hypotheses, a structured technique for keeping an agent (and the person reviewing its work) from anchoring on one explanation too early

• ​Collection requirements, which force you to define what an agent actually needs to know before it starts acting

​"I, Sage Faraday, ran an intelligence analysis shop as a Master Sergeant. I now use these same methods in consulting work with companies putting agents into production, and they hold up. You don't need a clearance or an engineering background to apply them."

Talk followed by Q&A.