Zero. That’s the number of dollars publicly attached to Gamma’s acquisition of Lica, the Accel-backed design startup whose co-founders will now run a new AI design research lab inside Gamma. Terms were undisclosed, which tells us nothing about the price and quite a lot about the intent. Companies disclose numbers when the number is the story. Here, the story is the lab.
I want to argue that this deal is less about presentation software than it looks, and more about a problem that everyone building agent systems runs into eventually: agents have gotten reasonably good at reasoning and remarkably bad at handing their work back to humans.
The last mile of agent output
Most agent architectures today are optimized end-to-end for text. A model plans in text, calls tools with text, receives observations as text, and emits text. That uniformity is what made tool-calling agents tractable in the first place — one representation, one loss function, one interface. The cost is that the final artifact, the thing a human actually reads and decides on, inherits the shape of the intermediate representation. You get a wall of prose, or a bulleted flattening of prose, and the human does the work of turning it into understanding.
Lica’s product was interesting precisely because it attacked that boundary from the other side. It took screenshots and recordings — dense, unstructured, visually-encoded traces of what someone actually did — and turned them into presentations and videos. That’s an inversion of the usual pipeline. Instead of text in, visuals out, it was visual evidence in, structured narrative out.
For anyone working on agent architecture, that direction should be familiar. It’s the same problem as building a computer-use agent that can watch a screen and infer intent. It’s the same problem as summarizing a long tool-execution trace into something a reviewer can audit. Perception-to-narrative is a general capability that happens to have been packaged as a design tool.
Why a research lab and not a feature team
The structural choice here is the interesting part. Gamma did not absorb Lica’s product into a roadmap item. It stood up a research division and put the acquired founders in charge of it, with a stated focus on new presentation methods. Reporting around the deal frames the goal as moving past conventional slide-based communication and toward output that adapts to its audience.
Read that as an architectural bet rather than a marketing line. “Adapts to its audience” implies the artifact is no longer a static file. It implies a generation policy that conditions on who is receiving the information, what they already know, how much time they have, and what decision they’re trying to make. In agent terms, the presentation stops being a document and becomes a rendering of a model’s internal state under audience-specific constraints.
That is a research question, not an engineering ticket. A few reasons it resists ordinary product work:
- No good objective function. We can measure whether a summary is factually consistent. We have much weaker tools for measuring whether a visual explanation actually transferred understanding to a specific person.
- The output space is structured and compositional. Layout, hierarchy, motion, and pacing interact. Sampling tokens left to right is a poor fit for artifacts where a late choice constrains an early one.
- Evaluation requires humans in the loop. Which means slow iteration cycles, which means you need a team whose job is the cycle rather than the ship date.
What I’d watch for technically
If this lab produces anything worth reading, I’d expect it to show up in one of three places. First, intermediate representations: some scene graph or document IR that a model can plan over before anything gets rendered, analogous to how code agents plan over an AST rather than over characters. Second, audience conditioning: an explicit model of the recipient, which is really a user model, which is really the same infrastructure personalization and memory systems need. Third, evaluation: benchmarks for communicative effectiveness that go beyond aesthetic preference ratings. That third one would be the most useful contribution to the field and the least likely to be prioritized.
The broader pattern
Acquisitions like this one are becoming a recognizable move — buy a small team with unusual taste in a narrow modality, then hand them a research mandate instead of a feature backlog. It works when the acquired team’s problem turns out to be a general one wearing a specific costume. Lica’s costume was presentations. The problem underneath is how machine reasoning becomes human understanding, and that problem sits directly on top of every agent system anyone is currently shipping.
Undisclosed terms mean we can’t judge whether Gamma paid a sensible price. What we can judge, eventually, is whether a design research lab produces ideas that other people’s agents end up borrowing. That’s a higher bar than a better slide generator, and it’s the one worth holding them to.
🕒 Published: