If you work in VC, you spend a disproportionate amount of time zooming in. You look at customer cohorts, question a growth spike, debate whether gross margins are sustainable, and spend hours trying to understand whether a startup is building a feature or a company that could eventually be worth billions.
Last weekend, Lorenzo, Irene, Tommaso and I were lucky enough to be invited to the Santa Fe Institute in New Mexico for its annual symposium, which this year was dedicated to Agentic Technologies. We spent two days discussing thermodynamics, neuroscience, evolutionary theory, electricity grids, the nature of intelligence, and even beach volleyball.
For those not familiar with it, the Santa Fe Institute is one of the most fascinating research institutions in the world. Founded in 1984 by a group of scientists including George Cowan, David Pines and Nobel Prize-winning physicist Murray Gell-Mann, SFI emerged from a fairly radical idea: the most interesting scientific questions don’t fit neatly into academic disciplines.
The founders wanted an institution without traditional departmental boundaries, where physicists could work with biologists, economists with mathematicians, and computer scientists with researchers studying human behaviour. The idea was to study complexity: how interactions between relatively simple components can generate behaviours that are difficult, sometimes impossible, to predict by looking at each component individually.
Think about an ant colony, a city, a financial market, the human brain, or even a startup. Understanding each individual element doesn’t necessarily tell you how the whole system will behave.
Over the past forty years, SFI has helped develop some of the most influential frameworks in complexity science, from network theory to agent-based modelling and the study of complex economies.
And today, with AI systems becoming increasingly autonomous, many of the questions that SFI has been studying for decades feel incredibly relevant. The symposium brought together people working across physics, biology, neuroscience, economics, robotics and AI, alongside people building, deploying and investing in new technologies.
What I particularly enjoyed was that people didn’t necessarily agree. Sometimes they couldn’t even agree on the definition of the thing they were discussing (amazing, I know).
So rather than trying to summarize two days of discussions (which would probably require a small book), I wanted to share five ideas that stayed with me, and why I think they matter for founders and investors.
Subscribe for more articles like this!
1) We are building an entire industry around a word we haven't fully defined
Let’s start with the obvious one: agents.
Every other startup pitch I see today mentions AI agents. Agents for sales, customer support, procurement, healthcare, accounting, coding, recruiting. The list is endless.
But what exactly makes something an agent?
One of the examples used to introduce the symposium was a thermostat.
A thermostat measures temperature, compares it to a desired state and takes action. In a basic sense, it has a goal and a decision rule.
Does that make it an agent?
It depends who you ask.
One of the most interesting tensions throughout the discussions was between the different meanings of agency across disciplines.
In neuroscience, for example, behaviour that looks purposeful doesn’t necessarily imply conscious decision-making. Biological systems can perform incredibly sophisticated, context-sensitive actions without anything resembling deliberate choice.
In economics, agency often means something else entirely: the ability to act on someone else’s behalf.
And in AI, we tend to use the word for systems that can pursue goals, select actions and execute tasks with different degrees of autonomy.
Three concepts that I found particularly useful to distinguish were autonomy, agency and delegation:
Autonomy is whether something can act without real-time human intervention.
Agency concerns its ability to select actions in pursuit of an objective.
Delegation is about who gives that system decision-making authority, and within what limits.
These distinctions might sound academic, but I think they are becoming extremely important commercially.
Imagine an AI customer-support system. There is a big difference between one that recommends how to respond to a customer, one that independently sends the response, and one that can also issue refunds, modify orders and make commercial decisions on behalf of the business.
All three might be marketed as agents.
But their economic value, reliability requirements and potential consequences are very different.
The same applies to an AI procurement platform that moves from analysing supplier data to negotiating contracts and writing changes back into an ERP.
As software moves from recommending to deciding and executing, we’re delegating progressively more authority to machines.
And that introduces a whole set of questions that I think founders building in this space should be spending serious time on.
What decisions is the agent actually allowed to make? How does it know when to ask for approval? How can you verify its actions? Who is responsible when something goes wrong?
I increasingly think the defensibility of successful agentic companies will depend as much on their ability to integrate deeply into existing workflows, manage permissions and build trust as on the underlying intelligence of their models.
2) New technologies don't replace old ones. They build on top of them
One of the frameworks we found particularly interesting comes from the work of economist W. Brian Arthur, who has spent decades studying the relationship between technology and economic evolution.
We tend to describe technological revolutions as discrete events. First came the internet… Then cloud computing… Then deep learning… Then generative AI. Now agents.
Almost as if one technology arrives, replaces the previous one, and takes over. But technology doesn’t really evolve that way. New technologies are combinations of technologies that already exist.
Cloud computing became possible because of advances in computing, networking, virtualization and data centres.
Generative AI became possible because of the combination of neural network architectures, large datasets, GPUs and scalable computing infrastructure.
Agentic AI is now emerging on top of those foundations.
Each layer builds on the previous ones, and the technologies underneath become the infrastructure that enables the next generation.
The interesting part is what happens to the economy when these new capabilities become available. Arthur’s framework suggests that technology doesn’t simply improve the productivity of existing industries. It progressively creates and recreates the structure of the economy itself.
The internet didn’t just make communication faster. It enabled entirely new business models, global supply chains, digital marketplaces and industries that previously couldn’t exist. Cloud computing didn’t just make servers cheaper. It changed who could build software, how companies were financed and how quickly products could reach customers. So what does agentic AI do?
I think we’re still extremely early in understanding the answer.
But here’s one way I find useful to think about it:
A traditional CRM stores information about customers. A more sophisticated CRM helps a marketer analyse that information and execute campaigns.
An AI-native CRM can increasingly understand customer conversations, decide whom to contact, determine the right message and execute the action (Yes, I am looking at TextYess!).
When a new technology layer emerges, it’s tempting to focus on the companies building the most impressive applications on top of it.
But I think the more important question is: which parts of the new stack will become critical infrastructure for the next generation? Will it be the underlying models? The orchestration layer? Proprietary data? Distribution? Deep integrations into existing systems? The application that owns the customer relationship? Probably different answers in different industries. And the answer will change as the underlying technologies become cheaper and more widely available.
One question I’m increasingly asking founders is: if the models you use today become ten times better and ten times cheaper, which part of your company becomes more valuable?
I think it’s a useful way to distinguish companies benefiting from a technological shift from companies that could become part of the next layer of infrastructure themselves.
3) One of AI's biggest bottlenecks might have very little to do with AI
Now, let’s move from software to electricity (Yes, electricity). One of the sessions I found particularly relevant explored the power grid as a complex adaptive system. We’re experiencing several transformations simultaneously:
AI data centres are creating enormous new electricity demand
Industries are electrifying
Electric vehicles are becoming more common
Renewable energy is increasingly distributed
Weather patterns are changing.
And all of this is happening on top of electricity infrastructure that was largely designed for a very different world.
The fundamental challenge is that these systems are deeply interconnected. An apparently sensible decision made by one participant can create problems when thousands of other participants make the same decision.
Take electric vehicles. Imagine a utility introduces cheaper electricity after 10pm to encourage people to charge their cars outside peak hours. Makes sense, right? Now imagine everyone programmes their EV to start charging at exactly 10pm. Congratulations, you’ve created a new demand peak. Every individual car is behaving rationally according to its instructions. The collective outcome is not necessarily good for the system.
This is one of the central ideas of complexity science: optimizing individual components doesn’t necessarily optimize the system.
There is another problem: time.
A new AI data centre can potentially be deployed in around 18 months. Building the electricity infrastructure required to support it can take several years, sometimes considerably longer. You can improve an AI model in a few months. You cannot improve the construction timeline of a major transmission project in the same way. And no amount of better prompting will change the laws of physics.
For investors, this creates a very interesting set of opportunities beyond the obvious AI software and infrastructure companies. Think about energy storage, distributed generation, grid automation, power electronics, demand response, data-centre energy management and technologies that allow compute capacity to operate more flexibly. (stay tuned for investments in this space!)
The winners might not simply be the companies that produce more electricity or build more data centres. They could also be the companies that make the entire system work better together.
There is also an important lesson for founders. If you’re building a company that interacts with the physical world, the constraints of that world are part of your product. The more AI moves into robotics, manufacturing, energy and healthcare, the more important that distinction becomes.
4) The best team isn't necessarily the sum of the best individuals
One of the most unexpected connections to venture came from a research presentation about sports.
The question was simple: how do individual contributions translate into team performance?
We often assume that if you assemble a group of highly talented individuals, you should get a highly performing team. But that’s not necessarily how collective performance works. Different activities depend on their participants in fundamentally different ways. In some systems, performance is disproportionately driven by the strongest participant. In others, performance is constrained by the weakest participant. And in others, what matters is a relatively balanced contribution from everyone.
The research explored this through different sports, including basketball and beach volleyball.
In basketball, particularly at the highest professional levels, a superstar can have an outsized impact on the team’s performance. You can give your best player the ball, design plays around them and concentrate opportunity where it creates the most value. Basketball therefore tends towards what researchers call a strongest-link structure.
Beach volleyball sits much closer to the opposite end. An opponent can deliberately target the weaker player, and both teammates need to participate successfully in maintaining a rally. In that environment, the weakest link can have a disproportionately large influence on the team’s outcome.
These are patterns identified in particular datasets, not universal laws about either sport. But the underlying framework is fascinating.
Because it applies to organizations too.
One of the things we kept discussing is whether AI is turning startups from an additive game into a disjunctive one.Historically, building a startup was largely an additive game. You needed engineers to build, designers to design, salespeople to sell. The more talented people you assembled, the more you could collectively achieve. But AI is changing that equation. As AI raises everyone’s baseline capabilities, a small team can increasingly accomplish what previously required dozens of people. And suddenly, the difference between a good team and an exceptional one may depend less on the sum of everyone’s contributions and much more on the outsized impact of one or two extraordinary individuals.
In other words, startups might increasingly start looking like basketball: a game where your strongest player can make a disproportionate difference.
The paradox? The more AI democratizes execution, the more valuable exceptional talent, judgment and creativity might become.
Of course, this doesn’t apply equally to every function. As companies scale, sales, retention, operations and compliance can still behave like weakest-link systems. But at the earliest stages, I wonder whether we’re moving from a world where having ten good people mattered most to one where having two exceptional ones matters even more.
And that has pretty interesting implications for how we invest in founders, build teams and think about talent in an AI-native world.
Again, think about a very early-stage startup. Sometimes a single exceptional technical founder can make an enormous difference. One person can build something that would otherwise require a team of ten. But when that company starts scaling, the bottlenecks change. An exceptional product team doesn’t compensate for a sales organization that can’t convert demand. An exceptional sales team doesn’t compensate for a product customers don’t retain. And in highly regulated industries, one unreliable component can compromise the entire operation.
The same question becomes even more interesting when we start thinking about organizations made up of AI agents. If you have ten agents working on a task, does adding a more intelligent agent improve the system? Not necessarily. If every agent must complete its task correctly for the final output to be useful, the system might be limited by the least reliable agent or handoff.
But if the task is structured so that the best agent can solve most of the problem independently, the dynamics look very different.
I think this has important implications for how we build teams, design workflows and evaluate AI products.
We spend a lot of time measuring individual performance.
Perhaps we should spend more time understanding how individual performance aggregates.
5) The most interesting opportunities might be the ones we cannot predict yet
This last idea is probably the one I have spent the most time thinking about since coming back. It emerged from discussions around assembly theory, evolution and the nature of creativity.
At a very high level, assembly theory explores how complex things emerge from the combination of simpler things, and how their existence depends on the history of what came before them.
But one of the questions raised was particularly interesting: what’s the difference between predicting something that could happen and creating something that was never possible before?
We often use probability and possibility almost interchangeably.
But they are not the same thing (!)
You can assign probabilities to outcomes within a space you already understand. If you know the possible outcomes of a process, you can estimate how likely each one is.
But what happens when a new possibility emerges?
Before airplanes existed, people could imagine flying, of course. But they couldn’t meaningfully forecast the economics of commercial aviation, airport infrastructure or global travel as we understand them today.
Those possibilities became economically meaningful only as new technologies, institutions and behaviours were assembled.
The philosophical argument is that evolution and creativity don’t simply explore a predefined set of possibilities. They can expand the space of what is possible. Whether this becomes a general scientific theory remains an open question. But I find the distinction incredibly useful for venture capital.
We spend a lot of time trying to predict the future. But some of the most interesting founders are actively creating a future that doesn’t exist yet.
Quoting the speaker: “Ultimate manifestation of intelligence is when an entity moves from dealing with probability to assembling possibility”
And that is inherently difficult to model. It’s relatively straightforward to analyse a company that makes an existing activity 30% cheaper or twice as fast. It’s much harder to underwrite a company building something that could fundamentally change what customers are able to do.
Are we investing in something that optimizes an activity we already understand? Or in something that makes a previously impossible, unaffordable or impractical activity possible?
Both can create great businesses.
But the second category is much harder to capture in a market-sizing spreadsheet, especially at the very beginning. Which is, of course, part of why venture capital exists.
One final thought: what happens to our own ability to think?
There was another theme running through several conversations.
If AI makes it dramatically easier to execute intellectual tasks, what happens to the skills we would otherwise have developed by doing them?
Imagine a PhD student who can now run sophisticated computational experiments in days instead of months. That’s extraordinary. But how does that student develop the intuition needed to recognize when the experiment is asking the wrong question? Or an engineer who can generate thousands of lines of code without writing them manually. How do they learn to identify the subtle mistakes that only become obvious after years of building things themselves?
I think there’s a very direct parallel with investing.
We’re increasingly able to use AI to summarize datarooms, analyse financial models, research markets and generate investment memos (all incredibly useful YES). But is the real value of an investor the ability to summarize information? Well.. not really. I think there is 10x more value in knowing which information matters, which assumptions are wrong, which questions haven’t been asked, and when a seemingly irrational founder might actually be seeing something everyone else is missing. “Models keep track of data, humans keep track of stories” > subtle difference in our job.
If anything, as the cost of producing answers approaches zero, the ability to ask good questions becomes more valuable.
And I suspect the same will be true for founders. Building products will become cheaper. Experimentation will become faster. More people will have access to capabilities that previously required specialized teams. But deciding what is worth building, understanding complex systems and recognizing opportunities that don’t yet look obvious will remain exceptionally difficult. “Providing answers is different than providing understanding”
That’s why we came back from Santa Fe feeling that spending time outside our usual conversations is not a distraction from the job. Two days surrounded by people studying everything from the behaviour of cells to the evolution of technology were a very useful reminder that the companies we’re evaluating are part of systems that are changing much faster than our investment frameworks.
And sometimes you need to get far enough away from the details to notice that the whole picture is changing.
The best founders I’ve met tend to be exceptionally good at doing both: zooming out far enough to see what is becoming possible, and zooming in obsessively enough to actually build it.
I think the same should be true for investors.
A huge thank you to the Santa Fe Institute for having us and for creating the kind of environment where very different disciplines can challenge each other’s assumptions.
And to Lorenzo, Irene and Tommaso for sharing the experience.
Virgi
