Everyone is talking about context and context graphs and context layers. If your LinkedIn feed is anything like mine, you are being barraged with a constant stream of tHoUGhT LEaDerShIp on this topic, replete with overwrought platitudes, apocryphal anecdotes, and complex diagrams.
And it all makes sense, in the kind of shallow way. Yes, AI needs context. Yes, it needs to be organized and defined. Yes, we’ll need systems to do this!
But how context gets built and compounds is fuzzier. And color me skeptical that anyone really has it figured out — or that it looks like the last generation’s architectures with AI bolted on.
What I know for sure, though, is that a system that will really work will help you build context automatically over time as a result of using a product. We call these invisible emissions “Context Exhaust,” and it’s the center of what we’re doing at Hex.
Creating context, by accident
In the course of using software every day, we all emit signals that can be learned from. The words I’m typing at this very moment are “context”, even though I didn’t set out to “generate context” specifically.
And when using an agentic system in particular there’s another layer to the context signal – not just what you’re creating, but your path to get there. How you frame your question, how you respond and pushback to the agent, what you accept and edit — all more “exhaust” to be learned from.
As a practical example we see every day, when someone uses our analytics agent in Hex, they come in, ask a question, and guide the agent to how they want to see the answer. “Query dim_customers, but only for accounts where they’ve been active for over 30 days” or, “no we use the labels on fct_revenue_changes to filter that down” or “format all of my charts with this color pallet”, or “now publish this app and share it with the revenue leads” — and on and on.
They’re literally explaining how their data is structured, how it should be interpreted, how they want to visualize it, and who the insights should be communicated to. And all of this is valuable not just in that chat, but potentially in every subsequent interaction they and everyone else are going to have.
The compounding feedback loop
This Context Exhaust — when captured and refined into semantic models, guides, and memories — bootstraps the base of knowledge that’s required to create accurate and trusted answers for everyone in an organization. It starts a flywheel of context, making it so the next interaction is more accurate, which in turn can generate more context, and on and on.
In Hex, we're enabling this by allowing data teams to drive this compounding cycle:
- They can sync in external contact sources, like their dbt repo, to get things started.
- Experts can start asking data questions and commit to their context repo directly from their agent interactions.
- Teams can observe data agent responses and get Warnings for when context may not be accurate or sufficient.
- The Context Agent creates Suggestions based on identified issues that teams can act on to drive improvement.
People ask more questions, they provide more context exhaust, and the cycle continues on.
Learning, together
In the pantheon of buzzy concepts in the analytics market, the idea of being “data driven” is at the pinnacle. The concept of everyone being able to ask data questions is the ideal every organization strives for, and the dream every vendor sells.
But in reality, being “data driven” isn’t about individuals looking at charts in isolation — it's about the whole organism working and learning together. It's a collective effort. It’s iterative. It’s evolutionary. Your organization’s knowledge can change, grow, and morph, constantly fed by the context exhaust captured on the edge.
The net effect is a system that gets smarter the more you use it — and an organization that does, too.
More on Data teams
