Whenever something becomes abundant, new products appear that were impossible while it was scarce.

That is an old rule, and the AI boom has created a great deal of brand new abundance. Magma found one piece of it that nobody was using.

What they built

Magma's pitch is four words: monetize your agent's traces.

A trace is the log of everything an AI agent did to finish a task. Which tools it called, in what order, what it tried, what failed, what worked. Developers produce these constantly for debugging and then delete them.

Magma's argument is that this material is worth money, because it is exactly what somebody training the next generation of agents needs. So a cost becomes something the company producing it can sell.

Ramtin Shahzanian is the founder. Its site is magmahq.ai.

Watching agents is already a busy corner. LangSmith, Langfuse, Braintrust and Helicone record agent runs so developers can debug them, with Arize and Datadog doing it at larger scale. On the buying side, Scale AI and Surge AI pay large sums for demonstration data, and Prime Intellect and Mechanize build the environments agents train in. Everyone records traces. Nobody treats them as inventory.

The point: abundance is what creates a new product

Look at what the new world has that the old world did not.

Three years ago, almost nobody was running software that thinks for itself in a long chain of steps. Today an enormous number of companies are, and every single run leaves a detailed record behind. That record did not exist in any quantity before, and now it is being produced by the million and thrown away by the million.

That is the shape that makes a business. Not a clever piece of technology, but a thing that has suddenly become plentiful and that nobody has worked out how to use yet.

Whenever there is a huge abundance of something new, there are new products to be sold. The abundance comes first and the product follows, which is why you find these by looking at what has changed rather than at what is hard.

It has happened many times. Cheap storage made photo sharing possible. Cheap mobile data made streaming possible. Cheap GPS in every pocket made ride hailing possible. In each case the new abundance arrived first, and the products people now think of as obvious were built on top of it afterwards.

Founders assume the AI boom is not for them

Here is the part worth carrying away, and it is about who gets to play.

No young founder seriously believes they will capture the skyrocketing demand at the centre of AI. That belongs to Anthropic, OpenAI, xAI, Google DeepMind, Nvidia and a handful of others. Training frontier models and making the chips needs capital and scale a small team will never have.

So most founders conclude the boom is something to build on top of rather than something they can take a piece of, and they go and build another application.

That conclusion is too quick. The giants are capturing the demand. Nobody is capturing the by-products, and a boom this size throws off enormous quantities of them.

You cannot compete with a giant for the thing everybody wants. You can quite easily be first to the thing the giant's customers are producing by accident and throwing in the bin.

That is what Magma found. Not a way to compete with the labs, but a resource created by everyone else's use of them, sitting unclaimed because it was filed under cost rather than under inventory.

It will happen again, in something nobody has named yet

This is my opinion, not something the company has said.

Magma is one instance of a pattern that is going to repeat, and the useful exercise is asking what else has just become abundant.

Plenty of candidates are already in plain sight. The prompts people write, in enormous volume, telling you what humans actually want from machines. Failed and rejected generations, which teach as much as successful ones and are deleted faster. Agent-to-agent traffic, which barely existed last year. Evaluation results, produced constantly by everyone and shared by nobody. Idle time on GPUs that were bought for peaks. Even the waste heat coming off the data centres, which is a physical abundance nobody planned for.

Each of those is currently somebody's cost line. At least one of them is somebody's next company.

The test is simple enough to run on a Sunday afternoon. What is now being produced in vast quantity that was rare three years ago? Who is throwing it away? And who would pay for it if it were collected properly?

Why it is bigger than it sounds

The short description is monetising agent logs. The method behind it is the valuable half.

  • Abundance creates the product. Cheap storage made photo sharing, cheap data made streaming. The new plentiful thing comes first, and the business follows.
  • The giants take the demand, not the by-products. A small team cannot compete for what everyone wants, and can easily be first to what everyone is binning.
  • The pattern repeats. Prompts, failed generations, agent-to-agent traffic, evaluation results, idle GPUs. Each is a cost line today and at least one is a company.

What to watch

The measure is whether any lab has actually paid for traces sourced this way. It is a young market and plenty of data ideas never find a buyer at the price the seller imagined.

The second is who owns the trace. If the company running the agent can sell it without asking its own customers, that is a problem waiting to happen, and the answer to that question matters more than the technology.

The third is whether Magma stays on traces or goes after the next abundance once it has the machinery for buying, cleaning and selling this kind of material. That machinery is the real asset, and traces are only the first thing to put through it.