Trajectory demos one-click training for agents

Sequoia Capital hosted Ronak Malde to demo Trajectory, a one-click train interface for agents. Malde said the platform turns agent traces into deployed models that run six times cheaper and lift accuracy from 4% to 18% for open source baselines. No attacker model or vulnerability was demoed. The affected systems are enterprise agent stacks that log traces for training.

Before this, teams stitched training, eval and deployment across separate tools. Then Trajectory put those steps in one place with support for GRPO and DeepSeek v4.1 flash.

The building block is GRPO. The method samples a group of answers per prompt and scores each one against its group average, so it needs no value model, according to DEV Community. Trajectory applies that approach to customer traces. Harvey and Rogo were named as early users. Rogo was cited as several times more token efficient and cheaper and faster after training on Trajectory, while Nvidia was also listed as a customer.

Malde, Trajectory founder, framed the bet as continual learning for every company. "My goal is to basically take that 1% and turn that into 99% so that we are actually learning from all of this built-up experience that lives in the world right now," he said.

The limit is data. The loop only closes if teams can define "what good looks like" and collect clean traces. Malde said only 1% of enterprises have adopted AI, but Sequoia Capital provided no methodology for that figure and Trajectory disclosed no independent benchmark for the 4% to 18% claim.

What still has to happen is the loop itself. Malde said Trajectory will let teams upload a trained model back to production, gather new traces, and retrain. No public launch window or filing date was given for that continual cycle.