Bringing Continual Learning to Every Company: Trajectory’s Ronak Malde
Ronak Malde, co-founder of Trajectory, argues that agents need to learn continuously from real-world experience instead of relying on offline batch updates. He covers how training custom models has become accessible to companies, two research directions (Group Size-1 GRPO and a meta-grader for reward propagation), the Harvey and Rogo results, the distinction between continual learning and recursive self-improvement, and a vision of product builders as gardeners who shape products that grow on their own.
Continual learning as the missing loop between intelligence and experience
Today’s agents are intelligent but updated in offline batches, closer to a paper road map than a live navigation system. The argument here is that closing the loop, so that deployed models learn from real-world traces, is the next learning paradigm, and that the tooling to do so is finally moving out of frontier labs and into ordinary companies.
From pseudo continual learning to the real thing
The analogy offered is AAA paper maps versus Google Maps. Today’s agents resemble the paper map: highly capable, but rerouting means a full batch update. The near-term goal is MapQuest-style democratization, where any company can train and customize a model. The end state is Google Maps-style learning, where every trip, or in this case every trajectory, updates the model in aggregate. Human cognition is the reference point for what the real thing looks like.
On architecture, the speaker does not think continual learning requires waiting for a new paradigm. Recent base models already incorporate changes such as fast and slow memory within the weights and skip connections across distant layers. The gap is less about the base architecture and more about connecting those capabilities to how models are actually used in the real world, and translating a company’s sense of “what good looks like” into concrete model improvements.
Why training used to be out of reach
Training a frontier model historically required a large research team, hundreds of millions in R&D and compute, and six to nine months of work, usually with consulting-style handholding. The speaker’s experience at Windsurf, where production edits and retries were used as learning signal, convinced them that the pattern was repeatable. Their co-founders had seen similar alignment problems from robotics and multimodal work at other labs. The insight was that frontier labs distribute one massive model to everyone, whereas customization for a single customer requires infrastructure that is currently only built inside those labs.
Harvey is the working example. Built on closed frontier models, it sat on decades of domain expertise but could not turn that data into an advantage. With an expert-built legal benchmark and the ability to train, they took an open-source baseline from roughly 4% to 18% accuracy, and produced a model about six times cheaper than the frontier, which allows far more tokens to be spent on the same task. Rogo followed a similar pattern, optimizing for intelligence density: a model that is more token-efficient, cheaper, and faster, while improving on the metrics the customer cares about.
Why the unit of work is a trajectory
The company name reflects a view that tokens are the wrong unit of economically meaningful work. In RL, a trajectory is a rollout, and in continual learning the rollout engine becomes live inference in production. Each trajectory then represents a task of economic value. The framing also gives companies ownership of their own learning path.
The pipeline: traces, replay, and reward
The stack is described in three parts:
- Traces. Companies need full observability over production runs, including sub-agent calls, intermediate steps, and user feedback. Existing observability vendors cover this layer.
- Replay. A trace must be replayable as an environment, which in practical terms means something like a Dockerfile or system requirements. Building faithful environments is hard, and perfect replication is often impossible.
- Modification. This is the training stack itself: loss functions, reward functions, and the algorithm. This is where in-house work is most difficult, and it is the layer the product aims to abstract away.
Two research bets: group size one and the meta-grader
Standard GRPO requires a group of 8 to 16 rollouts of the same task in the same environment to compare outcomes. Humans do not learn that way; they experience the world and act once. The proposed Group Size-1 variant learns from single real-world rollouts drawn from different tasks, which removes the need for a simulated environment and much of the infrastructure around it.
The reward problem is labeling. Legal and investment banking workflows rarely have a clean verifiable reward. The approach is to select about 100 representative traces from a pool of 10,000, have the customer write dense rubrics or criteria for those, and train a meta-grader that propagates that judgment across the full set. The longer-term aspiration is an observability interface where a user circles the five things an agent does well and the five it does poorly, then clicks train.
Curiosity and the transition function
LLMs are strong at moving from state to next state, but they do not model the transition itself: what happens between holding a ball and throwing it. Humans constantly probe the transition functions of unfamiliar parts of the world, which builds the heuristics that make us capable. Quantifying curiosity as a measure of what the agent does not yet know about transitions, and having it act to resolve that uncertainty, is presented as a path to autonomous learning.
Continual learning versus recursive self-improvement
Both RSI and continual learning concern how agents learn from experience, but they differ in where humans sit. RSI puts a model in a black box to improve itself as an AI researcher, which raises obvious alignment concerns. Continual learning learns from interaction with humans and their tasks, and is designed to keep humans in the loop. The speaker sees alignment as something built into the learning algorithm itself.
Products as gardens
As harnesses become most of the product in AI-native companies, and as both the model and the harness self-improve, the product builder’s role shifts. The builder sets trellises, meaning the directions in which a product should grow, and the product then develops largely on its own. Taste is expressed by shaping where the company grows, rather than by specifying every feature.
Abundance through specialization
In the speaker’s most optimistic view, Adam Smith’s argument from The Wealth of Nations applies: abundance requires specialization. Today’s AI landscape resembles a world where everyone consults one oracle. The future is one where every customer can customize the model and harness for their own domain. Only about 1% of enterprises have meaningfully adopted AI, and the goal is to turn that into 99% so that the accumulated experience embedded in real workflows feeds back into intelligence.