Robert Ta introduces Clarity, an open-source harness for AI agents that ingests artifacts like interview transcripts and generates story tickets. Humans annotate these tickets to capture subjective perspectives, and those annotations persist across sessions so agents can replay runs with new focus and improve over time.
So if you're a product manager and you're thinking about how do I make sure as I use my agents, the next time I use the agent, it gets better and better. And if you're doing multiplayer, you're working with many other PMs. How do you make sure that every run that one of you executes with an agent, you annotate it and that compounds the multiplayer execution as well? What you're seeing here is the Clarity open source harness that we're going to be releasing soon for AI builders. AI product managers, AI engineers to use in your regular flow of work. You'll just have a skill slash Clarity init harness. And as you use Cloud Code or Codex or Copilot, whatever, your agents register into this Clarity harness. So we have a concept of an agent registry and your Cloud Code or whatever coding agent is going to spin up sessions. You're going to write stories, you're going to upload transcripts and stuff like that. So in this case, I uploaded some transcripts from a particular interview with a customer and then it pulled out a sample artifact of a story, I could actually see the flow of work that this agent took. Had a tool call ingesting the source packet, which was a transcript. It extracted story tickets for learning, and then a model call to update a story or to make a new story, and then proposing a diff. What I could do is I could be like, hey, I want to also focus more on hiring managers at Media Enterprise. So I want it to be a bit more specific. And then that's gone ahead and marked there, and I have my annotation there, that annotation would capture from my profile. And that way it has my subjective perspective taking of what would make a better ticket. What I could then do is replay it, and I could see it rerunning through and focusing now more on Medium Enterprise. That's one way we're thinking about how can we make it easy for an expert in the flow of work of managing agents and subagents, reviewing their runs, annotating them. So we connect the human to the agent run, and then that annotation goes goes into the curated context layer that then feeds better agent runs in the future and compounds, and multiplayer context compounding.
