Ronak Malde is co-founder and chief executive of Trajectory, which builds infrastructure for continual learning: helping AI agents improve from the corrections, tool calls, and failures they encounter after deployment.
Malde studied computer science at Stanford, completing a master’s degree with an AI concentration, and met future co-founders Michael Elabd and Arjun Karanam there. At Codeium, he helped ship Windsurf and trained SWE-1, an agentic coding model. He subsequently joined Google DeepMind, working on Antigravity and Gemini 3 before leaving to start Trajectory.
Founded in 2026, Trajectory launched with $15 million in financing and announced partnerships with Clay, Harvey, Decagon, Mercor, and Rogo. Its platform converts production agent traces into updates to models and their surrounding tools, prompts, and orchestration, with evaluation, approval, and deployment controls.
Learning from deployed behavior. Malde considers real user interactions more representative of deployed agents’ challenges than curated benchmarks. Corrections, successful tool calls, and failed workflows expose domain-specific signals that conventional training pipelines often discard.
On-policy self-distillation. His continual-learning research trains a model against a better-informed version of itself. Comparing their token-level probability distributions yields denser feedback from individual trajectories than assigning one reward to an entire response.
Training agents across long workflows. Extended tool-calling sequences can cause teacher and student behavior to diverge, producing repetitive hedging or malformed tool calls. Divergence-weighted training adjusts the influence of individual steps, allowing useful portions of a trajectory to remain trainable.
Preventing hint leakage. Privileged information can teach a model shortcuts unavailable in production. Malde proposes filtering hints into realistic guidance and using residual guidance to limit how strongly that information shifts the student’s behavior.
His wider research interests include how readily open models respond to post-training and behavioral adaptation and how neural-memory architectures improve long-context behavior. Fully continuous improvement remains an unresolved engineering challenge: existing systems still rely partly on offline updates, redeployment, and infrastructure that reconciles many simultaneous interactions.
Learning from production requires more than collecting traces. On-policy self-distillation supplies dense feedback, but long tool-calling trajectories expose divergence, hint leakage, and deployment challenges.
Why build harder benchmarks while production keeps generating experience?
