Every enterprise has a website. Within a few years, every enterprise will have a model.

Not a subscription to somebody else’s model. Their own fine-tuned and reinforcement-learned model on their own data, their own workflows, their own corrections.

The reason is not cost. A model shaped by your operational history is the one asset OpenAI and Anthropic cannot take from you. Anything you build on top of a frontier model, the next release can absorb. The weights you trained on your own record, it cannot.

That is the first of three bets Humpback AI is making. It is contrarian today. We think it will be obvious in three years.

The Record Is the Asset

The second bet is where the company gets concrete.

Reviewing and steering the work of agents becomes the largest single share of human labour in the enterprise. Today that is true only in software, where engineers spend their days reading what a model wrote. Soon it is true of every function. And it will not be one person at a desk; it will be reviewed across teams, the way a pull request is.

Now notice what a review actually is. A human looked at what an agent did and corrected it. That is a labelled trace. Millions of them, accumulating daily, describing exactly how your organisation makes decisions.

The record of agents acting, humans correcting, is the proprietary data. Not the documents sitting in the lake; those are everyone’s documents, written before the agents existed. The corrections are yours alone, and they feed straight back into the first bet.

The obstacle is that today you can inspect that record only at a sampled rate. Reading every trace, from every agent, for training and feedback and security and governance, at a cost that does not exceed the value of the work being reviewed, nobody has built that.

Agents Need Their Own Cloud

Every cloud built to date assumes a human as the primary user. A person opens an account, chooses an instance size, reads a bill. Agents will open their own accounts, deploy what they need, learn continuously from what they produce, and publish what they build. Sandboxes are the first sign of this and nowhere near enough. Identity, authorisation, and billing all need rebuilding when the customer is not a person.

Put the three bets together, and this is not a product. It is a rebuild of the cloud from the account model upward — where an agent runs, how much infrastructure it commands, what authority it carries, and what record it leaves behind to train the next model.

Humpback’s own framing is that this is as large as the build-out that created AWS, and that it has barely started. We agree, and we think the window to be early is measured in quarters.

Why Prudhvi

A thesis this large is only as good as the person chasing it.

At Rubrik, Prudhvi Dharmana spent four and a half years on backup at petabyte scale. The hard part was never the backup. It was that you cannot test a system against a filesystem containing a billion files, because you cannot obtain one to test against. So he helped build one in memory, then wrote the engineering series explaining how. It is still on Rubrik’s blog under his name.

At TruEra, he was an engineering manager building AI observability for enterprises, at a company founded out of CMU that Snowflake acquired in 2024. He then spent two more years at Snowflake on agent monitoring, agentic evaluation, and optimisation.

Read as a list, that is a solid infrastructure career. Read as a sequence, it is one question asked three times, each time a layer higher: how do you know a system you cannot see into is doing what you think it is doing? For filesystems, the answer was simulation. For models, observability. For agents, nobody has an answer yet — which is exactly the gap the second bet describes.

He has been walking toward this problem for a decade, mostly without planning to.

What They Refuse to Build

The sharpest line in Humpback’s thesis is not about what they will build. It is about what they will not: anything that exists because a human needed it. Consoles. Dashboards. Catalogues of machine sizes. Abstractions whose only job is to make infrastructure legible to a person.

That is harder than it sounds, because those artefacts are precisely what enterprise buyers ask for in the first meeting, and shipping them is the easiest way to close the first cheque. A company holding this line will lose deals it could have won.

We think holding it is the whole bet. If agents genuinely become the primary users of infrastructure, every abstraction built for human eyes becomes dead weight, and the companies carrying it will spend a decade migrating away from their own products.

Proud to Partner with Humpback AI

Humpback is building the layer agents will run on. We’re excited to back Prudhvi on this journey.