Certain companies are pioneers, defining their categories before anyone else can see them. They also tend to do it more than once. At Trajectory, we work with a handful of them, and right now, they are all looking at the same thing on the horizon.

AI is more capable than ever, and yet it is still frozen the moment it ships. Users now do real work on top of these models and products. All of that work -- every retry, every correction, every edit -- is signal that the model could learn from. And almost all of it dies by morning. The model and the product are separated, and improvement only flows one way, from model releases into products, never back the other direction.

However, a different kind of product is coming into view. One where the model gets better at the work for as long as the work continues. One where the people who understand the domain best are the ones shaping what the model becomes. One that grows with the people using it instead of waiting for the next release. The technology that’ll underpin this is called continual learning, and we’re in the earliest innings of what it’ll become.

But the path won’t be easy. The playbook does not exist, and there are fundamental research, infrastructure, and product questions that have yet to be answered. These teams have started to build for this future despite that, in production, with real users, in real domains.

Today we are highlighting a few of these pioneers.

———————————

Harvey

Legal expertise lives at the long tail of human knowledge. It is deeply specialized and continually shaped by precedent, judgment, and practice in ways fixed datasets cannot capture. With Trajectory, we are building toward agents that can carry the evolving domain expertise behind this important work. - Gabe Pereyra, President & Co-Founder of Harvey

Harvey is pushing the frontier of what legal AI can do. They are setting themselves a high bar, in a domain where eighty percent right is not eighty percent useful. To make that standard concrete, they recently open-sourced a benchmark for legal agents, graded by expert rubrics that mirror how partners actually review work. With Trajectory, they are building on that signal toward models that continually improve.

Clay

"Continual learning is an important research direction for Clay's roadmap, and Trajectory is building the infrastructure to help us explore it," said Kareem Amin, CEO of Clay. "We're testing a model that is getting smarter over time from our users, we've already seen examples of it learning from its mistakes.”

Clay is what AI-native GTM looks like when it is built from first principles. They have spent years thinking about what the work actually requires, the primitives, the patterns, the ways operators compose their own approach to it. Now they are working with Trajectory on how models can learn that craft, and then keep getting better at it over time.

Decagon

"We deploy AI across enterprise customers where the right model behavior shifts meaningfully from one setting to the next, which makes model steerability not just a research interest but a core operational requirement. Our collaboration with Trajectory is focused on understanding how to rigorously measure steerability in post-trained models, and what training techniques actually improve it. Steerability is often claimed but rarely characterized precisely, and we're trying to build the diagnostic and training foundations to change that. That is what it takes to let a system of specialized models keep improving in production." - Cyrus Asgari, Research Engineer at Decagon

Decagon is what enterprise AI support looks like when the models powering it are built for the work, not borrowed from somewhere else. Over eighty percent of their traffic now runs on models they trained themselves. With that architecture comes a harder problem: each enterprise has its own definition of the right answer, shaped by their customers, their tone, their policies, the specific things they cannot get wrong. Holding all of that in a model is what makes the whole system work. With Trajectory, they are working on how a system of models like this can keep learning in production, without breaking the behaviors their customers depend on.

Mercor

Creating datasets that actually improve models has historically been more art than science. Trajectory's platform has allowed us to incorporate post-training further up in the data production lifecycle, validating training signal before data goes to customer researchers. Through our collaboration with Trajectory, we have no technical barriers to continuous validation of training signal for new datasets as we produce them. - Michael Haines, Product Lead at Mercor

Mercor is what the expert layer underneath frontier AI looks like. They were early to a structural insight about modern AI: what separates a capable model from a useful one is the quality of the human judgment going into training, and at scale. To make that judgment measurable, they have built benchmarks that grade AI agents against the standard of real professional work, with experts from law, banking, consulting, and other domains writing the criteria. They are now working with Trajectory on what that signal can become inside the training loop itself.

What's next?

Trajectory is a research and product lab building the platform for continual learning.

Users shape how a model actually works through the corrections they make, the retries they run, the small moments where they push it toward what they actually meant. Trajectory is the interface where that gets turned into training signal for the next version of the model, across the whole stack: the model, the harness, and the prompt layer learn together. We are building it as a platform so that every research insight and every infrastructure advance can reach every team building real products, and empowering those teams to build the products only they can make.

More is coming. You can see our early work on the algorithms, infrastructure, and platform for continual learning. But the real evidence is in the work, and deep dives with each of these teams are coming soon.

Here is to the teams at the frontier, and to the ones who will join them there. We are in this together, building toward what comes next.