Riddle (August 2025)

Riddle was an in-feed "attention agent" launched on Farcaster. There are quite a few projects on Farcaster building tools to programatically buy engagement from other users in order to boost content on the feed. These make up some of the alternatives to advertising on a decentralized and programmable social network. The issue was that these tools incentivize any kind of raw engagement and not necessarily anything of quality.

Riddle was an AI agent that quality scores responses to a boosted piece of content in real time, creating a live incentivized contest design to drive quality engagement to a post that a user wants to highlight. We built a quality framework around three pillars: persuasiveness, creativity and authenticity. Within each pillar, we built semantic analysis heuristics intended to analyze language and convert those results into a quantitative scoring model.

The experiment generated a lot of engagement and quite a few learnings that we are looking forward to incorporating into future AI agents. As expected, many responses -- even by high authenticity users -- were LLM-generated, highlighting a common and growing challenge on social media. Our framework picked up on some common LLM patterns, but certainly not all of them. At a higher level, however, the exercise was more about scoring quality and less about the human element. Of course, humans prompting machines for content is an everyday authentic experience and will only become more mundane.