Last week brought us the AI Engineer World's Fair in San Francisco: 6,000+ engineers, founders, and VPs of AI packed into Moscone West for four days of talks, workshops, and expo-floor conversations. We were there too, talking to platform engineers about the problems they're actually hitting when they try to run agents in production.
A few patterns showed up often enough in those conversations that they're worth writing down before the sugar-high of a conference wears off and we forget half of it. Here’s an honest, open (that’s a theme here) take on being at the event.
1. We might have been the only open source company on the floor
This wasn't something we planned to lead with, but it kept becoming the conversation. Scan the sponsor list and the expo floor and its closed platforms, black-box agents, and a lot of "trust us" where you'd expect to see a GitHub link. When we told people our stack is open source, the reaction was genuine surprise, followed by real questions.
Our open source first approach seemed to really matter. It also opens the door to a joke that writes itself: our products aren't vibe-coded Python wrappers we have to keep out of public view... That's not a jab at anyone in particular, it's just an honest description of a category that's gotten crowded with things nobody can actually inspect. Maybe we should have leaned more into being the open, inspectable option, because right now almost nobody else in this space is making that claim credibly.
Key Takeaway: Trust is really important in a saturated, heavily vibe-coded market, with a lot of vendors hoping to ‘build the plane in the air’, on the customer’s dime.
2. Most people are early in the journey
A recurring pattern in booth conversations: a lot of attendees are still discovering that the problems we solve are problems at all. They're not yet at "how do I do granular cost attribution across 40 teams"... they're at "wait, this is a thing that happens to everyone?".
Because the problem someone has after two weeks of running agents in a dev cluster is a completely different problem than the one someone has after six months in production. Leading with the answer before you know the question wastes everyone’s time.
Key Takeaway: It isn’t reasonable to automatically jump straight to the most sophisticated version of the story in conversation with folks. The more advanced capabilities grabbing the attention of those at the tip of the spear aren’t where most are at today.
3. Two ideas resonated more than the rest
Two things came up over and over in one-on-one conversations, and they weren't the two things we necessarily expected to be the headliners:
- Virtual MCP and progressive disclosure. People immediately got why loading every tool's full schema into context doesn't scale, and the idea of exposing tools progressively (name and description up front, full schema on demand) landed as an obviously correct pattern rather than something we had to sell. The concept of progressive disclosure is easy to comprehend, but the awareness of simple solutions still needs work.
- Granular cost visibility. Close second, and probably the single best demo moment we had at the event. Showing per-user, per-team spend as a real, live visual (not a promised dashboard, an actual one, because we run our own products) turned heads in a way a slide or graphic never will.
Key Takeaway: While it's tempting to focus on what's next, attendees got the most value from conversations that started with the challenges they were facing today. Building from familiar problems made it easier to understand emerging approaches and technologies.
4. It's an engineer-focused event, but the room includes the people who own risk
AI Engineer World's Fair bills itself around engineers, and most of the floor traffic matches that. But founders and Heads of AI were there too, and their questions skewed differently… less "how does the config work" and more "how do we govern this," "who's accountable when an agent does something wrong," "what does this look like to a compliance team."
That's a distinct audience with a distinct set of concerns, and it's easy to miss if you're only tuned in to engineer-shaped questions.
Key Takeaway: Production AI deployments are still so new, and moving so fast, that people of all titles and experience are out to learn what they can, engineer or not.
The Bottom Line
None of this changes the plan, it merely sharpens it. We lean harder into open source as a genuine differentiator instead of a footnote. Meet people where they actually are in the problem, not where we assume they are. Keep pushing on the ideas that are already resonating instead of spreading the story thinner. And don't forget the founders and Heads of AI in the room just because the badge says "Engineer."
See everyone at the next one.
%20(1).png)
%20a%20Bad%20Idea.png)
