17 years verifying facts on the web.

“My biggest aha moment from the product.ai research report came when I read an important suggestion it made. It suggested enhancing the operational framework we have for context usage in our AI research repo by incorporating each block’s class and the tool calls it feeds into. That turned out to be the piece we were missing. In our first benchmark run, the structured approach read fewer files repeatedly but still spent more tokens overall, because it verified its work by re-reading everything it had just built. Declaring which tool each block feeds into means verification becomes running that tool and reading the result, instead of re-reading the block. And tagging each block with a class lets the agent skip whole categories of context it does not need for the current task, which is where most of our excess tokens were going. Two small metadata fields address both halves of the cost we measured.”