We got rid of our AI product team

Earlier this year we got rid of our Magic team, which was responsible for building AI features in Hex. This was a tough decision at the time, but over the last few months, has been roundly validated.

I want to share a bit about why we did this, what we learned, and some takeaways I believe any other company building in 2025 and beyond should internalize — especially those who have built dedicated "AI teams".

The dawn of Magic

Just over two years ago, we launched Hex Magic, a set of AI features meant to augment, not replace, data people. We were early (maybe the first?) folks in our space to do this, and the vibes and excitement were really great.

So naturally, we started building a dedicated Magic team of engineers, designers, and a PM, and surfed the AI wave forward.

This worked well for a while! Magic was a small, nimble team who could keep up with a fast-moving space and iterate quickly. By having one team building all our AI features, they could also make sure we were being consistent around design practices and think deeply about UX. They built new features, wasted time on some duds, and learned a lot of hard (even bitter) lessons on how to apply AI to analytics.

Coming into this year, however, something felt… off. Our Magicians were working hard, building a bunch of AI features, and coordinating with the rest of the team… but it wasn’t adding up.

Something had to change.

Too much Magic

We were seeing issues pop up in all sorts of places.

Magic features were effectively treated as bolt-ons. We’d build the mainline feature – say, a new way to do data viz – and then the Magic team would then think about how to wire it into our AI features. This meant that we often had to do re-work, and missed opportunities to approach features as truly AI-first.

This also meant that very few engineers in the company were actually eligible to work on AI features – most were plugging away on “mainline” work and Magic was treated as a completely separate group with some special, arcane knowledge that they weren’t allowed to be exposed to.

What started as a feature – that Magic had dedicated capacity for AI development – started to feel like a bug, as resources started to compete between these and "vanilla" features.

Over time this divide — between “Hex” and “Magic” — felt more and more glaring, and it was obvious that we weren’t operating from an AI-first mentality.

In fact, it reminded me a lot of the cloud 10 years ago – companies that treated it as an add-on or special extra thing were left behind, while those that internalized and built around it in a cloud-first way made it into the future.

As a young startup, I felt we had to make the decision on which side of that we wanted to be on – and we did, by disbanding Magic, and re-organizing into Persona and Platform teams.

Personas

At Hex, we have three distinct personas we care about – Editors (more technical), Explorers (less technical) and Curators (organizing and governing context). Our teams were already organized around this – so we decided to just collapse the engineers working on AI into those teams.

No more separate prioritization queues for AI vs. vanilla features – each team is now responsible for all of it, based on what’s going to move PMF for their target persona. AI feature? Great – build it. Non-AI thing? That works too, build that instead.

Now, given the massive potential of AI and how fun it is to work on, sure – teams have gravitated toward those things. But not exclusively! We’ve built a lot of great good-old-fashioned deterministic stuff this year, although even then the way we’ve approached it has changed. Now that more people are viewing the world through the lens of context and tools, it’s becoming more obvious where the interaction and integration points will live across these features.

This has literally 10x’d the number of people working on AI features – now everyone is an AI product engineer! Folks learned fast, and we’ve been impressed with how quickly people who have never touched AI have started building incredible new things.

This has been accelerated by rapid adoption of AI coding tools within our team (our team is split between Cursor and Claude Code) – which have quickly implanted mental models around tools and context into our team’s brains.

Building an AI platform

What we have kept is an AI platform team, which is responsible for core research, infrastructure, and frameworks. By having a team focused on these foundational capabilities, it lets our persona teams build on stable abstractions, move faster, and share components.

The team has also taken on a bunch of interesting things, including:

• Building fine-tuning and model-hosting infrastructure on top of Baseten and Modal, which has enabled us to deploy specialized models fine-tuned from open source models, as well as move to self-hosting all of our embedding models. We have also now fine-tuned our own typeahead model with great success (we'll share more in a future post!)

• Developing internal frameworks for building agentic products on top of Temporal for better resilience and observability

• Enhancing our internal eval infrastructure with tools for manual review as well as LLM-as-judge for agent threads

• Scaling and generalizing our internal RAG infrastructure backed by LanceDB to support agent context retrieval

Engineers on our AI Platform team are typically embedding with the Persona teams to focus on specific problems or pioneer new frameworks. As an example, two Platform folks worked full-time with Editor to build a new agentic framework to power the Notebook Agent – but now are focused on generalizing and abstracting that to power new features with other teams.

These boundaries can be pretty fluid – and we’re still figuring out exactly what should live where, when. But we’re confident this is the right direction of travel!

Going back to the cloud analogy – software companies today have cloud infra and devX teams that work on core abstractions for deployment, databases, and multiplayer – we think AI will move in a similar direction.

Magic is dead, long live Magic

So internally, and externally, standalone Magic is no longer treated as a separate team or sub-brand (even though it was extremely cool and led to some truly epic swag).

But Magic lives on, in that we are focused on making Hex feel magical because of how well we’ve incorporated AI into the bones of the platform – not because of some separate bolted-on AI thing. AI is the product!

This opens up some new, and exciting possibilities for us – many of which we’re launching in a few weeks.

I’m very excited about the future. It’s going to be magical.