Co-Scientists, Embedded Payments, and Constellation
multi-agent systems, vertical payments, and AI vs. Constellation
Google is out with a co-Scientist. And it turns out that they actually built a multi-agent system. The paper is fantastic, going as far as to give system construction, variable parameterization, and prompt construction. It notes the use of tools, web search, and more as crucial aspects of the system. And it also is careful to note that it’s built upon Gemini 2.0. Most of the specialized aspects of the system seem to be in specialized models such as AlphaFold.
This is no mere hardcoded workflow, it’s a dynamic system, elegant in its construction, and was capable of building multiple hypotheses that had not been publicly disclosed.
I don’t think startups have even scratched the surface around what’s possible with multi-agent systems. In fact, the current state of the art in vertical AI leans in the complete opposite direction: codify a workflow but utilize AI to handle randomness that may otherwise upset the workflow.
The workflow method today has crystal clear ROI: less busy work. And workflows most likely are not going away (we invent processes and workflows to gain control over randomness). But the prospect of multi-agent systems that actually lead to innovation/better outcomes is an underexplored direction in today’s market.
For instance, it seems somewhat obvious to me that based upon this paper, you could create a multi-agent co-tutor that could teach you or your children anything. Again, the system is elegant because it shows evidence that it can replicate across industries.
Second, it seems also obvious that businesses could benefit from true multi-agent systems for answering some of the trickiest strategic or analytical questions in their industries. So why is it still nascent?1
The tough things per usual are actually around the evaluation side. The reason we even know that this co-scientist works is because scientists had spent the last decade forming and proving the hypothesis that the co-scientist uncovered.
And I think this points to the fundamental constraint of multi-agent systems left to solve: feedback loops in the real world are not instantaneous. Things take time, operating in a complex world means that the decision may have been temporarily right, but long term wrong. Build a tutor system and you’re still waiting most likely years to measure outcomes. Craft a co-decisioning agent for businesses and it may be months or years before anyone knows if the system got it right.
This is again, why evals ultimately have to originate not in ethereal model capabilities but in real-world outcomes. Google just gave the first evidence that this approach is valuable by concealing a decade-long science research problem, continuing to build on this will be key.
Embedded Strategies are Continuing to Explode
AI is where everyone’s mindshare is, but let’s not ignore the continuing growth of embedded payments. Toast disclosed that they have now eclipsed $42 billion in payments volume, up 25% year over year. They additionally created $43 million in gross profits from their lending operations in Q4. Those are monstrous numbers.
The key question is whether this growth will continue across not only restaurants, but across every industry. I think it clearly will. What’s more, the vertical SaaS vendors are clearly benefitting from intense competition at the payfac level around payments. Adyen reports 28 embedded platforms processing over $1B in GPV per year. They further note that the North American market has become brutal with many solutions now competing on cost vs. experience.
Reading the tea leaves here, I’d expect that vertical SaaS vendors are finding ways to onboard payments that give them higher rev share and then also passing on some savings to merchants. This has always been the bull case for vertical SaaS. Make the payments experience cheaper, more integrated, and more accessible.
This may not be ideal for Stripe or Adyen, but it points to value creation continuing to happen at the vertical SaaS layer. And it still may be the early innings for not only payments, but for every other embedded initiative.
Will AI disrupt Constellation?
I can no longer find the tweet, but a very astute point was made the other day:
It goes something like this:
Constellation is able to reap value from vertical SaaS that is hard to displace and that customers never leave.
AI seems to imply that long-tail vertical markets will be easier to build solutions for.
These solutions by and large have low continued R&D expenditure. AI-native developers could theoretically create Constellation-type solutions for far cheaper.
This theoretically puts price pressure on Constellation companies.
Thus, if AI is actually going to put long term pressure on SaaS, we would expect it to show up in Constellation’s earnings and thus they represent a leading signal if and when this occurs.
I find this analysis fascinating. Your opinion is going to largely be based around a) distribution dynamics, b) the actual possibility of creating these long-tail software companies cheaply, and c) system of record driven moats.
But I think there’s one key reason why I’m unsure that this risk heavily impacts Constellation in the near term:
If a customer is already willing to use a highly legacy vertical software solution that’s well past its expiration date, shouldn’t we expect them to continue using it?
In other words, Constellation seems to specialize in industry applications that are important, but amongst a business subset that seems to not really care too much about being on the bleeding edge. Turns out, that’s a lot of businesses.
This slow play may make them one of the biggest winners of AI long term. After all, these long-tail AI companies are going to need exits. And who better to give them exits than Constellation? They may acquire their way to replacing their existing legacy install base on the timeline they think is best.
Should be fun to watch.
There is one obvious reason: reasoning agents just came out.