We build tools, not agents

When people learn that we run an applied lab, the assumption is usually that we are building an agent. We are not, and the reason is worth stating plainly because it determines everything else we do.

An agent, at its core, is a loop: a model, a set of tools it can call, and a cycle of deciding what to call next. That is a genuinely important invention and it is also, by now, a well understood one. The frontier laboratories ship agents, they ship them well, and they will keep shipping better ones on a cadence no small team can match. There is not much left to invent in the loop itself.

What there is a great deal left to invent in is everything the loop reaches for.

The plug already exists

The second half of the argument is the part that changed recently. For a while, connecting an agent to anything was bespoke work: every tool, every integration, its own adapter, its own conventions.

Then the Model Context Protocol arrived, introduced by Anthropic in November 2024. Within a year it had been adopted by OpenAI, Google and Microsoft, and in December 2025 it was handed to the Linux Foundation. A protocol that competitors adopt and that is then given to a neutral steward has stopped being a product decision and become infrastructure.

That matters for a small lab in a specific way. It means we do not have to win a platform fight to be reachable. If we build something genuinely useful and expose it properly, any agent can use it, including agents that do not exist yet, made by companies we will never talk to.

So the value moves into the tools

Put those two together and the conclusion is not subtle. The agents are commoditised and the connection is standardised, so what is left to build, and what is left to be valuable, is what the agents reach for.

The market is not small. Around 97 million monthly downloads of the protocol's software development kits as of March 2026. More than ten thousand public servers in production, reported by Anthropic in December 2025. Gartner expects 40 percent of enterprise applications to embed task-specific agents by the end of 2026, up from under 5 percent in 2025, and puts agentic spending at 202 billion dollars in 2026, up 141 percent in a year.

Those numbers describe demand for the loop. Our bet is that the durable position is one step behind it, in what the loop is allowed to do.

What a tool for agents has to be

Building for this reader changes the work in ways we did not anticipate.

A tool for agents needs a real surface, not a fenced-off subset of what the interface can do. Anything a person can accomplish by hand, an agent should be able to accomplish through the protocol, or one kind of user has quietly become second class and the workarounds start within weeks.

It needs to be legible. An agent cannot infer what a vague label means or squint at a half-loaded page, so the vocabulary has to be precise and the states explicit. That discipline improves the human side of the product too, which is a happy accident we have stopped being surprised by.

And it needs to belong to the user rather than to us. The agent a person works through is the one they chose and configured, the one that knows them. A product that insists you use its agent instead is fighting the direction people are already moving in, which is toward centralising on one agent and connecting it outward to everything.

Which is why the tools are open

If the agentic layer belongs to the user, then the tools it reaches should be inspectable by the user and, increasingly, by the agent itself. An agent working inside an open tool can read the source of the thing it is standing in, answer its own question about how a behaviour works, and file a fix rather than a complaint.

We publish what we build for that reason and for a simpler one. Code is cheap now, so owning code is no longer where the value sits. Owning the fit is: the work of making a tool right for one company's actual situation. The base can be open, and usually should be.

Three tools so far, each used inside our own Studio every day before anybody else sees it. That is the test we trust, and our first customers were our own agents.