AI is the most capable software ever built. It writes, it reasons, it codes.

And yet, as you use it, you find yourself nudging it. Refining. Correcting. Providing helpful context to improve it. But the next day you come back, and it knows none of this. Whatever you taught it is gone.

Nothing else intelligent works this way. Minds learn. Languages learn. Cities learn. Your immune system learns. Everything we recognize as intelligent gets shaped by what it encounters. AI is the exception. The smartest thing we have ever built is the one that does not get smarter from being used.

This is deeply unnatural, and it does not have to be this way.

A handful of products are starting to feel different. Claude Code. Cursor Composer. Windsurf SWE-1. In each one, the model and the product are no longer separate things. The team building also shapes the model. The model is being trained on how the product gets used. They move as one.

This is the beginning of continual learning, and we are building Trajectory to make it possible for everyone.

What it takes

Continual learning has to happen across three surfaces: the model weights, the harness around it, and the prompts that drive it. No one yet knows how to treat them as a single, unified system. Making this real takes two things, built in parallel.

Frontier Research. How do you turn the signal coming off a deployment into something a model can learn from? How do you stop treating the model, the harness, and the prompts as three separate systems, and start shaping them as one? And how do you build the infrastructure to make this continual, so intelligence compounds daily instead of yearly? These are the open questions. And they only get answered with scale and realism of data from production.

A Platform. Today's attempts at shaping model behavior are manual and ad hoc. Teams either post-train by hand, or are stuck playing prompt-whack-a-mole: capturing trajectories, deciding what to train on or what prompts to modify, pushing changes through one at a time. That works for one product. It doesn't scale. The platform is what makes continual learning something a team can do themselves. One where you optimize the model, the harness, and the prompts jointly, in one place.

In its final form, the platform will look more like a creative tool. You shape your intelligence the way a designer shapes a design: refining it until it matches your intent.

This is why Trajectory is a research lab and a product lab in one. The research only happens if you build the product. The product only works if you do the research.

To work, both must run on the same shared unit - something the research can turn into learning signal, and something the platform can let teams shape.

The primitive

That shared unit is the trajectory.

Every era of computing has been built off a core primitive. The row and the query gave us databases. The page and the link gave us the web. The request and the response gave us the cloud. Each defined what could be built on top of it.

The primitive of this era will be the trajectory: the trace of what the agent did, and the telemetry of what the user did with it. The trace is what the field has been training on. The telemetry is the part everyone has been throwing away. Which suggestions the user accepted, which they overrode, where they came back to fix something the model got wrong, these are the signals that tells you whether the model succeeded, and almost no one is using it.

A trace tells you what happened. Telemetry hints at whether it should have. A trajectory is both. And you need both to teach. We named the company after the primitive, because this primitive is our bet.

The plan

Our ambition is continual learning. But the path to demonstrating value does not start with a grand abstract system. It starts with the thing every AI product already needs: better performance.

Step 1: Make existing AI products cheaper, faster, and better.

Right now, companies need systems that are faster, cheaper, more reliable, and better at the work. We use the signal already coming off their products, the edits, retries, rejections, escalations, acceptances, and outcomes, to make models that are better, faster, and cheaper. This gives us the right to live in the production path.

Step 2: Give teams a control surface for shaping their intelligence.

Once inside production, we see where intelligence actually fails: missed intents, brittle tools, silent edits, prompt regressions, and behaviors teams wish they could change. We also see that the model is only part of the system. Intelligence lives across the model, prompts, harness, evals, and product itself. Today, those levers are scattered, making your intelligence difficult to control. Trajectory will be the tool that brings them into one place.

Step 3: True Continual Learning.

A control surface gives teams a way to shape the intelligence inside their own products. But the real prize is not shaping it once. It is helping their products keep improving because they are used. Every deployment should create more signal. Every signal should sharpen the next update. Every update should make their product more useful, more specific, and harder to copy. While everyone else gives you the brilliant PhD on day one, Trajectory helps turn that intelligence into someone with ten years on the job.

Put more simply? First make it better, then make it shapeable, then make it learn. The pieces are all there, and one path serves every company that runs on intelligence, from the smallest startup to the largest enterprise. That is how Trajectory becomes the bedrock of every product that comes next.

What comes next

For 70 years, software has been something engineers built and everyone else used, with a hard line between the two. You could file a ticket. Hope for a new release. But the software did not change because you used it.

We are building Trajectory to be the place to build a product that learns, the way the cloud became the obvious place to build a product that scales.

The era of frozen software is ending. Continual learning is what comes next.