Ronak Malde on Continual Learning from Real-World Agent Traces

Sequoia’s hosts ask Trajectory co-founder Ronak Malde what he means by continual learning. He compares today’s agents to paper road maps: useful, but updated in offline batches. His goal is a system closer to a live navigation service that learns from use and updates as conditions change.

Asked whether current architectures can support that goal, Malde says architectural changes are already appearing in foundation models and will matter. Trajectory’s chosen layer is to connect that research to real-world usage and translate a company’s idea of a good outcome into model improvements.

He says training a frontier model previously required large research teams, significant compute and months of work. At Windsurf, he worked on learning from edits and retries; his co-founders brought experience from robotics and multimodal models. He saw a recurring problem: turning real-world feedback into better models.

Large labs tend to train one model for broad distribution, he argues, whereas tailoring a model to each customer requires different infrastructure. He presents Trajectory as a way to make that customization accessible to companies without each building a full research stack.

Malde cites legal AI company Harvey as a customer. He says an open-source baseline rose from around 4% to 18% on Harvey’s legal-agent benchmark and that the resulting model was six times cheaper than a frontier alternative. These are company-reported figures on a particular benchmark, not a general accuracy comparison.

The name Trajectory reflects his preferred unit of work: a task or rollout, rather than a count of tokens. In his framing, production interactions can supply real-world trajectories from which a training system might learn, provided the team can define and measure what success means.

He identifies several ingredients. First, capture production traces: the main agent, subagents, product events and user feedback. Some companies build this observability themselves; others use external tools. A useful trace is a timeline of what happened, not just the final answer.

Second, make traces replayable for evaluation or training. That requires some representation of the environment and its dependencies. Third, build the training stack and reward or loss functions that translate behavior into model changes—work Malde wants the platform to abstract for customers.

Two hard problems remain: a replay environment may not faithfully reproduce production, and legal or finance work rarely has a perfectly verifiable reward. Malde describes two research directions intended to reduce the manual work in those areas.

He explains group relative policy optimization using a game analogy: conventional grouped rollouts compare several attempts at the same task. Trajectory is exploring a group-size-one variant that learns from single rollouts across different tasks. He says this can avoid constructing a repeated simulated environment for that approach; it is a research claim with its own assumptions.

For rewards, he proposes selecting roughly 100 representative traces from 10,000, having experts supply dense criteria or rubrics for those examples, and training a meta-grader to generalize across the larger set. He says it is not a perfect solution and describes a simpler observability-and-training interface as a future goal.

Malde says Trajectory’s released platform brings training, evaluation and deployment together: a customer supplies agent traces and an account of what good looks like, trains a model, evaluates it and deploys it. He envisions collecting new production traces and repeating the cycle.

The initial audience is AI-native companies that already have an agent harness. Malde also wants to improve harnesses, arguing that a customized model may improve efficiency on familiar tasks while tools and orchestration can enable new capabilities. His suggested routing split between frontier and custom models is a future scenario, not a measured guarantee.

He cites Rogo as another example, saying it wants a concise finance agent with greater “intelligence density.” He claims a model trained with Trajectory used fewer tokens and was cheaper and faster while improving on measures the customer cared about; the talk does not provide a full independent evaluation.

Looking further ahead, Malde says agents need a way to accumulate experience and explore what they do not know. He uses a child learning to throw a ball as an analogy for building a model of how actions change the world. He frames curiosity as a research problem rather than a capability already solved.

He describes state–action transitions in reinforcement learning and asks whether an agent could identify uncertain transitions, try tasks and learn from the results. This is his proposed direction for autonomous continual learning, with substantial technical and safety questions still open.

Asked about alignment, he distinguishes his vision of continual learning from a version of recursive self-improvement in which a system improves itself in isolation. He wants models to learn from interactions and tasks chosen by people, keeping humans in the loop. That distinction describes his approach; human involvement alone does not prove a system is aligned.

Malde says Trajectory combines frontier research and product engineering. The company has productionized parts of the training stack with services and autoscaling, while design engineers run training jobs to understand the researcher’s workflow. He says coding agents help the team spend more time discussing system design.

He imagines product builders becoming gardeners: they set boundaries and direction while models, harnesses and eventually products improve from experience. This is a forecast. He also says the broader ecosystem needs observability, sandboxes and other infrastructure beyond Trajectory itself.

In his optimistic view, specialized systems could connect the general intelligence of foundation models with experience accumulated in individual domains. He compares this to specialization in an economy and hopes customization will bring AI to more workflows. The adoption figures and predictions he offers are his own estimates.