Recombine, don't invent

This year I have watched two-person teams use agentic AI to create millions of pounds of value in a few months, and watched large companies spend millions on the same technology and get almost nothing back. The easy explanations are talent, agility, culture. I think the real one is more specific, and more useful: the two are making different kinds of move, and only one of them pays.

It is the same asymmetry my last two essays kept circling. The same agent that transforms one person's output becomes a net drain inside an organisation.¹ Here is the mechanism underneath it. The move that pays is recombination, and its strangest property is this: the move that is safest for the people winning right now is the single most dangerous move for the people losing.

Three ways to build something new

There are, roughly, three ways to innovate.²

You can invent a module: a genuinely new component, the first lidar sensor, a new model architecture, a new material. It is what most people picture when they hear "innovation," and it is slow, capital-hungry and fails often.

You can improve a module: swap one part of a system for a better one, a new engine in the same car. Safe and incremental, but only if you already own the car.

Or you can recombine: leave the components alone and wire them together in a configuration nobody had built. Existing film libraries, existing streaming and existing recommendation algorithms, combined, are Netflix. None of the parts were new. The arrangement was. Any cocktail is the same trick: the spirits already exist, and the drink is the proportions.

The textbook, written from the chair of a large firm, treats recombination as the dangerous one: reconfiguring how everything connects breaks the routines, the handoffs and the org chart, so incumbents stick to safe, incremental improvement of what they own. Recombination gets "left for new entrants" because only they are reckless enough to try it.

The data says the new entrants should send a thank-you note. When researchers classified 298,915 venture-funded startups by how they actually innovated, the ranking that holds for incumbents flipped completely.² The ventures that recombined existing modules reached IPOs and high-value acquisitions; they got from seed to early funding nearly three years faster and drew a more diverse set of investors. The ventures that tried to invent new modules took longer and failed more. In the authors' words, "what is typically seen as the riskiest form of innovation can, for startups, be the safest route to success."

What AI just changed

Agentic AI has done one specific thing to this picture: it has collapsed the cost of recombination.

The expensive part of recombining was never the idea; it was the labour of making separate modules actually talk to each other. That labour (holding context, calling tools, reading a schema, stitching systems together) is exactly what a capable agent is good at. So the move the data already marked as the safest, highest-return path for a new venture just became dramatically cheaper to execute, for anyone.

You can see it in the breakout companies built on the models. Cursor did not train a frontier model; it recombined a code editor, version control and someone else's model into one workflow, and became one of the fastest-growing software products on record. Perplexity recombined web search, a language model and inline citations into an answer engine. Harvey wired a frontier model to the documents and review workflows of a law firm. Even Claude Code and Devin are recombinations: a model joined to a terminal, a filesystem, version control and a test runner, parts that all already existed. The components were sitting in the economy; the arrangement, and the agent doing the wiring, were the new bit.

This is why the two-person team wins, and it has nothing to do with being cleverer. They own no legacy system, so improving a module isn't available; there is nothing to improve. Inventing one is too slow. What they can do, now cheaply, is recombine what exists into an arrangement nobody had bothered to build. Every incentive points them at the one move that works, and AI just dropped its price.

Why the same move sinks a large company

The large company is trapped on the other side of the same table, and its own prudence is the trap.

Its instinct, often hardened into governance, is the "safe" move: bolt the agent onto the existing process and change nothing else. For an organisation that already owns a production system, that yields almost nothing. You buy the licence, the work still flows through the unchanged process, and you now pay for the new thing while carrying the full cost of the old one.

The move that would pay is recombination: rebuilding the workflow around cheap, fast execution, keeping the modules that earn their place and re-wiring how they connect. But that is precisely the move that, for an incumbent, is most expensive, because it breaks the architecture the whole organisation is balanced on: the approval gates, the data locked in legacy systems, the handoffs, the ownership boundaries. So the large company defaults to the bolt-on that doesn't pay and avoids the recombination that does. "Play it safe" is not the cautious choice here. It is the losing one, and because the asymmetry is structural rather than a matter of effort, the gap is likely to persist rather than close on its own.

This trap is not new, and neither is the way out. Thirty years ago Clayton Christensen diagnosed almost exactly this in The Innovator's Dilemma: well-run incumbents fail at disruptive technology not through incompetence but because the processes and values that make them good at today's business starve the new thing of resources.³ His prescription was structural: hand the disruptive business to "an organization small enough to get excited about small opportunities" rather than forcing it through the parent's machinery. Read against the startup data, that is sharper than ever. A separate unit works for the same reason the two-person team does: no legacy architecture to break, so recombination is cheap for it too. IBM built its first PC in a deliberately independent shop in Boca Raton; HP let a separate division build the inkjet that would come to compete with its own laser business. The firms that get real value from agentic AI will often be the ones that refuse to route it through the centre.

Invention isn't useless, it's a different layer

None of this makes inventing new modules a mistake. It puts invention one layer down, on a longer clock. Someone has to invent the module before anyone can recombine it, and that work is slow, costly and failure-prone, exactly as the data warns. But once a new module standardises, it stops being a risky invention and becomes cheap raw material for everyone else. GPT-3 cost a fortune to train in 2020; within five years open-weight models of comparable capability (Llama, Mistral, DeepSeek) had pushed the price of that module towards zero. Today's invention is tomorrow's recombinable part.

There is a design lesson in that. Invention only becomes a reusable module if it is standardised: clean interfaces, legible schemas, explicit contracts. A pipeline only your team can operate is a cost that never becomes an asset; the same work, built to be recombined by others, becomes infrastructure. It is why some of the most consequential releases of the past year have been standards, not products: the Model Context Protocol turns one team's integration work into a connector any agent can plug into. That is the difference between paving a public road and laying a private driveway. When you pay the invention price, build the module, not the monument.

What to do with this

• Stop rewarding the safe bolt-on. An agent in an unchanged process feels prudent and returns nothing; the upside is in reconfiguration, which feels riskier and isn't.
• Make recombination the core question: not "where can we put an agent?" but "what does this workflow look like rebuilt around cheap, fast execution?"
• If you can't recombine the parent, spin out a unit with no parent to protect. Christensen's prescription still holds.
• Buy recombination; reserve invention for where you have an edge. If a module already exists, recombining it is the fast, fundable path.
• When you invent, build modules, not monuments: standardise it so the rest of the organisation can recombine it.

A fair objection is that the biggest prizes of the AI era have gone to inventors: Nvidia became the first $5-trillion company in October 2025, and the leading model-builders are valued many times larger than the largest recombiners.⁴ But that is the high-variance end of the table, not a refutation of it. Behind the prizes sits a graveyard of near-collapses and fire-sales, exactly the tail the startup data predicts for module-invention. Recombination is the other end: smaller wins, far more of them, compounding faster on a fraction of the capital. By 2025 enterprises were already spending more on AI applications than on the models beneath them.⁴

For the overwhelming majority of organisations, the ones who will never train a frontier model, the choice was never invention in the first place. The ones creating value from AI aren't those with the best models; they're the ones recombining what already exists, faster than anyone burdened with a system they have to protect. The strategic question isn't which new thing should we build. It is can we recombine our own house before someone smaller recombines it for us.