In this presentation, Robert Ta introduces the open‑source Clarity harness, a stacked, persona‑specific AI framework designed for product managers. The harness combines multiple specialized agents—story writer, interview capture, requirements synthesizer, and an orchestrator—to ingest stakeholder transcripts, extract and refine requirements, generate user stories, and propose code changes. Human reviewers can annotate, approve, or adjust agent outputs, preserving provenance that links each decision back to the original stakeholder input. By aggregating these annotated agent runs across dozens or hundreds of product managers, the system creates a curated, queryable knowledge base that ties learning to concrete business outcomes such as adoption metrics and code quality. Ta emphasizes that this stacked harness architecture enables organizations to harvest accelerated learning from frontline workers, make data‑driven decisions, and continuously improve AI‑assisted product development.
Chapters
If you're a product manager, this is how we're evolving our work, I should say. What I'm about to show you is our own Clarity harness that we're about to ship open source. And this harness takes the philosophy that every individual has their own individual harness. There should be domain-specific harnesses. There should also be department or company-wide harnesses. The whole thinking is if I'm a product manager, let's just take that persona with this particular case. I'm going to have like a story writer agent at the top. We have this Clarity orchestrator agent. I have this interview agent that captures input. I have like a requirements synthesizer agent. So I could like open these up in this harness and see a view where I could actually very easily see, oh, this was from this transcript. Hey, we should think about a way to scale this as services stuff. We pulled this— my story writer agent worked with the requirements synthesizer agent to pull from the transcript what are the requirements and how do we write this as a story. And then you as the judge could be like, okay, this was quoted from here, What was the chain that got me here? This is the transcript. What did we add? What was the diff from before? What's the full artifact look like? Cool. That's the requirement. I could approve it. I could annotate it. I could say this requirement was too narrow. We should better specify broader. And I could save that and that saves in my Harness and then I could regenerate it. The idea is your agents that you're working with, like if you're working with Cloud Code, it's gonna spin up subagents. If you're working with your own agents, you're going to point it to say, hey, like use the skill, orchestrator calls the agent, calls the skill kind of thing. So with a harness like this where you're in the flow of work, you're writing your stories, you're making your prototypes, blah, blah, blah, you have agents doing stuff. The idea is there are agent runs and you want to know, did the input that I put in get me the expected output? And then how do I make that even better next time? That's what you're doing. And then if you want to do that across 100 PMs, 200 PMs, 300 PMs, this is what it takes is you get those product managers managers to be able to annotate the agent runs. And I could see here, I could see, okay, we ingested transcript, we extracted requirements, the model made a decision, new requirement or updating an existing requirement. Cool. We didn't take the other path. There's a new requirement. We propose a diff and now it's waiting for me as a human to review. So this is how we're, we're dogfooding this open source harness for clarity as we get requirements from our own customers acting as forward deployed engineers and AI founders. And we go in here and we're trying to annotate these stories and requirements, et cetera. So that our harnesses compound so that as I talk to more prospects, more customers, all of it could be curated learning that is queryable. Even as the team grows, we have the provenance of why we made these decisions in the product down to first-party data, down to this stakeholder, this customer, this prospect said this thing, and hey, Robert the human made this judgment call that this was the right requirement. This is how we formed the acceptance criteria for this story, and that informed the input spec to do this code change for this pull request and hey, did that increase adoption? Did that move the needle on our business goals? And that's like the individual harness moves to agent harnesses moves to is the code good? So domain-specific codebase harness to now business outcomes. And that's how we see it is you need to have a stacked harness architecture if you want to harvest all of your accelerated learning from the edges of the org, the frontline workers, the PMs, the designers, the engineers that are learning and implementing every day. They're learning and now using agents to do it. And how do you get all those agents and people way better decision-making. Make all the data queryable, make all the learning queryable, make it curated, and tie it to real evidence.
