For much of the past two decades, I have argued that the architecture of the enterprise needed to change.
The traditional enterprise was built around integration. Large applications, tightly coupled processes, centralized data and organizational structures designed primarily around functional specialization. This architecture made considerable sense when the principal objective was efficiency at scale. But it came with a cost. As enterprises became more complex, changing them became progressively harder. Dependencies accumulated. Systems calcified. Eventually, even relatively modest changes could require coordinating technology, process, data and organizational changes across dozens of boundaries.
The answer, I argued, was the Composable Enterprise.[1]
The underlying idea was straightforward. Instead of constructing the enterprise as a collection of large, tightly coupled systems, we should decompose it into smaller, modular capabilities with well-defined interfaces. Those capabilities could then be assembled and reassembled in response to changing circumstances. In the original 2013 formulation, I described matching Component Operating and Component Architecture Models—COM and CAM—which together allowed business and technology capabilities to be treated as reusable components and ultimately as an “Enterprise as a Service.”[1]
Much of what has happened in enterprise technology since has moved in this direction. Cloud computing, APIs, microservices, containers, data products and modular platforms are different manifestations of the same architectural principle. They reduce the cost of recombination.
But I increasingly think composability was only the first part of the story.
The arrival of generative AI, and particularly the emerging world of agents capable of using tools, invoking services and collaborating with other agents, suggests that something more fundamental may now be possible.[2]
We may be moving from enterprises that are designed to change to enterprises in which change itself increases the capacity for further change.
I think of this as the Autocatalytic Enterprise.
A lesson from four billion years ago
To explain what I mean, it helps to go back rather further than the history of information technology.
Something extraordinary happened on Earth roughly four billion years ago.
There was chemistry, but no biology. Molecules interacted with other molecules according to the laws of physics and chemistry. At some point, however, collections of chemical reactions became organized in ways that could sustain increasingly complex structures and, eventually, the processes we recognize as life.
Exactly how this happened remains one of science’s great unanswered questions. There are several competing and overlapping hypotheses about the origin of life, and we should be careful not to turn any one of them into established fact.[3]
One of the most interesting ideas comes from the study of autocatalytic systems.
Stuart Kauffman, whose work on complexity has influenced my thinking for many years, explored the possibility that sufficiently diverse collections of molecules could form networks in which the products of some reactions catalyzed other reactions within the same network.[3][4] Kauffman’s early work on collectively autocatalytic sets was subsequently developed into the mathematical framework now known as RAF—Reflexively Autocatalytic and Food-generated—theory.
The important idea isn’t that a single molecule somehow reproduced itself. It is that a network of relationships became collectively self-reinforcing.
A helps produce B. B enables C. C makes the production of A easier.
At some point the network ceases to be merely a collection of independent reactions. The products of the system contribute to the system’s ability to produce more products.
That is autocatalysis.
And while the analogy should not be stretched too far, I think something remarkably similar can happen with innovation.
When innovation produces the capacity for more innovation
We normally think about innovation in terms of outputs.
A new product. A new process. A new business model. A new technology.
We invest resources and, if things go well, an innovation emerges.
But some innovations are different.
They don’t merely produce an output. They increase the capacity of the system to produce subsequent innovations.
Consider the printing press.
Its immediate innovation was cheaper and more scalable reproduction of text. But its larger effect was recursive. Cheaper books distributed knowledge more widely. Wider access to knowledge allowed more people to build upon existing ideas. Those people produced new knowledge, which created more material worth printing and distributing.
The output became part of the input.
The scientific revolution produced similar dynamics. A scientific discovery frequently produces not simply knowledge, but instruments, methods, mathematical techniques and theories that make additional discoveries possible.
Industrialization provides another example. Better machine tools allowed the production of better machines, which enabled the manufacture of still better machine tools.
Semiconductors may be one of the clearest modern examples. More powerful computers enabled increasingly sophisticated semiconductor design and manufacturing. Those processes produced more powerful computers, which could then be used to design the next generation.
Software accelerated the process again. Programming languages produced libraries. Libraries became frameworks. Frameworks were exposed through APIs. Open-source communities created enormous repositories of reusable capability. Each generation reduced the amount of work required to construct the next.
These systems differ enormously in their mechanics, economics and institutional structures. I am not suggesting they are manifestations of some universal law.
But they share an interesting property.
Innovation becomes autocatalytic when the products of innovation increase the system’s capacity to produce subsequent innovations.
That distinction matters.
What composability was really buying us
Seen through this lens, the significance of the Composable Enterprise looks somewhat different.
We generally justified composability in terms of agility.
Break large systems into smaller components. Standardize their interfaces. Reduce dependencies. Make capabilities reusable. Allow teams to assemble solutions rather than repeatedly constructing them from scratch.[1]
All of that remains true.
But the deeper benefit of composability may be that it creates the substrate from which autocatalytic innovation can emerge.
A monolithic enterprise has relatively few possible combinations of capability. Creating something new often requires changing the underlying system itself.
A composable enterprise has many more possibilities.
A customer capability can be combined with a payments capability, a location capability and a pricing capability to produce something that wasn’t anticipated when any of them was originally created.
The value of each component therefore isn’t limited to the purpose for which it was built. Part of its value lies in the future combinations it makes possible.
This is an important distinction.
Composability increases the enterprise’s capacity to change. It does not, by itself, create the propensity to change.
A company can have beautiful APIs, modular applications, reusable data products and a sophisticated cloud architecture and remain spectacularly bad at innovation.
Something else has to happen.
The components need to interact.
The results of those interactions need to be captured.
Successful combinations need to become reusable capabilities themselves.
And those new capabilities need to expand the space of possible combinations available to everyone else.
Once that feedback loop exists, something qualitatively different begins to emerge.
Composability turns the enterprise into a collection of building blocks.
Autocatalysis turns those building blocks into an innovation system.
And then the components became intelligent
This is where AI changes the trajectory.
Until recently, most enterprise components were passive.
An API didn’t decide when it should be invoked. A database didn’t go looking for a problem to solve. A microservice didn’t discover another microservice and decide that combining their capabilities might produce something useful.
People did that.
Humans remained the principal agents of recombination.
We discovered the problem, identified the relevant capabilities, understood their interfaces, wrote the software, evaluated the result and decided what should happen next.
That constraint is beginning to disappear.
AI systems can increasingly interpret context, reason about objectives, discover tools, generate software, invoke services, analyze outcomes and use those outcomes to determine subsequent actions.[2]
The components of the composable enterprise are acquiring agency.
This doesn’t mean that applications have suddenly become autonomous organisms, nor does it require the breathless assumption that humans are about to disappear from the enterprise.
It means something much more practical.
The cost of discovering and creating useful combinations of capabilities is falling.
And that matters because the number of possible combinations inside a sufficiently large composable system is enormous.
We have spent twenty years creating the pieces.
We are now beginning to create machines capable of helping us discover what the pieces can do together.
From composable to autocatalytic
That suggests an evolution in how we think about enterprise architecture.
The monolithic enterprise was optimized primarily for integration and efficiency.
The composable enterprise added modularity and recombination, dramatically increasing adaptability.
The autocatalytic enterprise adds feedback: new combinations create new capabilities, which themselves become available for subsequent combinations.
AI adds another dimension because the participants in the system are no longer entirely passive. Increasingly intelligent components can participate in discovering, constructing and evaluating those combinations.
This is why I am reluctant to describe what is happening simply as another wave of automation.
Automation substitutes technology for an existing activity.
Autocatalysis changes the productive capacity of the system itself.
Imagine an organization in which an agent develops a better way to reconcile a particular class of transactions. That solution isn’t simply executed. It becomes a reusable capability available to other agents and teams.
Another agent subsequently discovers that the reconciliation capability can solve part of a different problem.
That combination generates another capability.
Its results improve a model.
The improved model increases the performance of several other processes.
Those improvements generate better data.
The better data enables still more capable models.
No individual step is particularly mysterious.
What matters is the topology of the whole system.
Capability produces capability.
The metric we’ve been missing
This also suggests that we may be measuring innovation incorrectly.
Most organizations measure innovation as output.
How many new products did we launch? How many experiments did we conduct? How many AI applications did we deploy? How much revenue came from products introduced during the last three years?
Those are reasonable measures.
But they don’t tell us whether the organization’s capacity to innovate is itself increasing.
That may be the more interesting variable.
If an organization conducts one hundred experiments this year instead of fifty last year, we might reasonably conclude that it is innovating more.
But suppose the cost of conducting the 101st experiment is essentially identical to the cost of conducting the first.
Compare that with an organization where every experiment leaves behind reusable data, code, models, knowledge and capabilities that make the next experiment cheaper and faster.
These are fundamentally different systems.
One is producing innovation.
The other is producing innovative capacity.
In mathematical terms, we have traditionally been preoccupied with I: the quantity of innovation.
Perhaps we should be paying considerably more attention to dI/dt: whether the system’s capacity for innovation is increasing over time.
That should eventually be measurable.
Does the time from idea to experiment decline?
Does the marginal cost of experimentation fall?
What proportion of new capabilities reuse existing capabilities?
How frequently does something created for one purpose become an input into something its creators never anticipated?
How quickly does learning in one part of the organization propagate into others?
And, perhaps most importantly, does each generation of innovation expand or constrain the space available to the next?
These are different questions from the ones we usually ask about digital transformation.
Rethinking the AI business case
They also lead to a different way of thinking about AI investment.
Most companies currently approach AI with a perfectly understandable question:
Where can AI make us more productive?
Where can we reduce cost? Automate work? Increase throughput? Improve customer service? Make employees more efficient?
There is nothing wrong with those questions. There is an enormous amount of value available from answering them.
But they are first-order questions.
The more interesting question may be:
Which investments increase our capacity to create the next generation of capabilities?
That produces a very different investment calculus.
An AI application that automates a specific task might produce an excellent conventional ROI while contributing almost nothing to the organization’s future innovative capacity.
Another investment—in reusable data, models, agents, interfaces, knowledge or experimentation infrastructure—might initially appear less compelling when evaluated against a single use case.
But that isn’t necessarily its economic boundary.
Its output becomes someone else’s input.
The resulting capability becomes another component.
That component participates in combinations that weren’t contemplated in the original business case.
Some of those combinations create still more reusable capabilities.
The investment begins to compound.
This is where conventional ROI analysis becomes increasingly uncomfortable, because the most consequential returns may occur several generations away from the original investment.
The value of a capability is no longer simply what it does.
It is also what it makes possible next.
The Autocatalytic Enterprise
There is a temptation whenever a new technology arrives to declare everything that preceded it obsolete.
I think that gets this transition exactly backwards.
AI doesn’t make the Composable Enterprise irrelevant.
It makes composability considerably more important.
Intelligence embedded in a monolithic architecture remains constrained by the architecture. An agent that cannot discover capabilities, access data, invoke services or act across organizational boundaries may be intelligent, but it has remarkably little leverage.
Composability provides the substrate.
Intelligence provides increasingly capable actors within that substrate.
Feedback allows the products of their activity to become new components of the system.
Put those things together and we get the possibility of something we haven’t really had before: an enterprise in which innovation systematically increases the capacity for further innovation.
Not autonomous.
Not self-aware.
And certainly not some corporate organism magically liberated from human management.
Autocatalytic.
There is an important difference.
Four billion years ago, somewhere in the enormously complicated transition between chemistry and biology, collections of relatively simple components began participating in networks whose products contributed to the production of further products.[3][4]
We still don’t know precisely how that transition occurred, and autocatalytic-set theory is only one part of a much larger scientific conversation about the origins of life.
But the systems idea is profound.
Complex systems can cross thresholds at which their outputs begin expanding their capacity to produce future outputs.
For decades, we have been decomposing enterprises into smaller, reusable and increasingly interoperable components.
We called the resulting architecture composable.
Now we are adding intelligence to those components and connecting them through increasingly rich networks of data, models, agents and feedback.
The interesting question is no longer simply whether those components can be recombined.
It is whether their interactions can produce capabilities that increase the system’s capacity to produce still more capabilities.
The Composable Enterprise was designed to change.
The Autocatalytic Enterprise compounds its capacity to change.
That may turn out to be a much more consequential distinction than simply whether an enterprise has adopted AI.
References
[1] Jonathan Murray, “The Composable Enterprise,” Adamalthus, April 4, 2013.
This is the original formulation of the Composable Enterprise, including the Component Operating Model (COM), Component Architecture Model (CAM), reusable atomic business and technology components, and “Enterprise as a Service.”
The Composable Enterprise — Adamalthus
[2] OpenAI, “Introducing the Agents API,” September 10, 2026; OpenAI API Reference, “Agents.”
Current examples of agent architectures incorporating tools, command execution, persistent context and subagent coordination.
Introducing the Agents API — OpenAI
Agents API Reference — OpenAI
[3] Wim Hordijk and Mike Steel, “Autocatalytic Networks at the Basis of Life’s Origin and Organization,” Life 8(4), 62, 2018.
A useful review of Kauffman’s collectively autocatalytic sets, their relationship to other models of minimal life, and their later mathematical formalization as RAF theory. The authors characterize autocatalytic sets as necessary but not sufficient for life-like behavior.
Full paper — Life
[4] Wim Hordijk and Mike Steel, “Chasing the tail: The emergence of autocatalytic networks,” BioSystems 152, 2017.
Traces autocatalytic-network theory from Kauffman’s work beginning in the early 1970s through later theoretical, computational and experimental research.
PubMed record
Related: Jonathan Murray, “Your Company Is Buying AI. Is It Learning How to Think?”, Adamalthus, August 14, 2026.
This earlier essay explicitly reconnects the 2013 Composable Enterprise model to AI-era capability compounding, including COM, CAM and Enterprise as a Service. It therefore provides the immediate intellectual bridge between the 2013 article and the Autocatalytic Enterprise argument developed here.
Your Company Is Buying AI. Is It Learning How to Think? — Adamalthus
