Robert Ta describes building a company brain using Clarity, an open-source harness that captures organizational learning through modular agents and digital customer twins. Every annotation and agent run gets stored in a data warehouse with full lineage, letting teams scale from thousands to tens of thousands of runs by feeding richer context back into the system.
Everybody wants this company brain thing, this elusive, like, how do we build a brain where all the edges of the enterprises are people, are humans? They're learning, our designers, researchers, product managers, sales, marketing. They're learning from prospects, from customers, from product analytics, synthesizing the results of experiments. And that way we take that learning and then we feed the next agent runs at scale. And one easy way to think about this is if we do 1,000 agent runs today across 1,000 people, let's say, then how do we make sure it's 10,000 tomorrow and those 10,000 are like 10x better because they have way better context, way better truth. That just fuzzy paper napkin math to help with that. The way we think about company brains as Clarity and what we're releasing with the Clarity open source harness is I am in the flow of work. I go, I'm using my PM Storywriter agent here and maybe any other agents I might need, like a research agent or design agent, or hey, I have these personas of customers or users in my ecosystem or my SKUs, and I might use this PM Storywriter agent. That's what I'll do. like initialize the harness from Cloud Code, and I'll throw into it, hey, there's like a transcript, and I want you to go ahead and take that out, extract the learnings we have from talking to customers, propose to me differences in like— in this particular case, it would be modeling somebody that I've talked to. That's what we do at Clarity is we model the self or make digital twins of customers, of stakeholders, of you. Mm-hmm. And that way, for example, when I annotate this. I might say, okay, product context learning may inform PM agent dogfood traces. So this was an internal run, and this is for us dogfooding Clarity. So I might say something like, this should expand to more than just— it should help with all agents and capture that. I could even say things like— so I could see what evidence this came from and what references this came from. I could see the quotes from a particular transcript and how that attributes here, and then I could replay it, right? And on the back end here, What we then save is those runs, the lineage of when I annotated this. So every annotation is a first-class citizen in your data warehouse now, right? That's what you should think about when you think about company brains is annotations, agent runs, those should be first-class citizens in terms of my data. And then from there, the question then becomes, then how do you connect this All of this so that it's multiplayer. But you could very easily see how, okay, cool. The next time I use this PM Storywriter agent, it will have that annotation. And underneath that leads to things like a harness getting better. This is one example of that. We'll manage your harness versions here. And this is the system prompt. Here's the grading rubric. Here's the failure taxonomy. Here's the dataset splits, skills and tools that the agent might be given. So all of these things should be treated as first-class citizens and things you should version and evaluate because this informs the quality bar of your agent over time. And that's especially important when you're thinking about delivering the right context at the right time to all parts of, of your company, AKA company brain.
