Valued at $1.15 billion, former DeepMind team founds Trajectory to address AI's feedback loop shortfall

According to Dynamic Beating monitoring, the continuous learning platform Trajectory, founded by former senior researchers from Google DeepMind, Apple, OpenAI, and Meta Superintelligence Labs, has officially debuted. The company simultaneously released the manifesto "Continual Learning: End of Frozen Software," advocating that AI software should not remain static and frozen after deployment. Instead, it should transform every user correction, retry, and coverage behavior into a direct reward signal, enabling online continuous updates of the model.

At the core of Trajectory is the stitching together of the agent's execution traces and the end-user telemetry. Execution traces accurately record the model's tool invocations and intermediate reasoning steps, while telemetry captures whether the user accepts, modifies, or backtracks on the results. The development team emphasizes that dynamic user intervention data on the output, typically discarded in traditional development, is, in fact, the golden training signal to assess whether the agent has truly mastered complex tasks.

On the technical implementation front, Trajectory integrates with mainstream observability platforms (such as LangSmith) through a standardized SDK. The toolkit automatically imports the agent's runtime trajectory, converting dialogue history, tool invocations, reward feedback, custom metrics, and error information into a standard Trajectory format for subsequent fine-tuning and evaluation. The accompanying web platform is currently in the development stage and will offer visual retrieval and refined management of trajectory data in the future.

In terms of commercialization and ecosystem development, Trajectory has completed a $15 million seed round, valuing the company at $115 million. This funding round was led by Conviction, with participation from Bessemer Venture Partners, Radical VC, and BoxGroup, and attracted investments from prominent individual investors such as Google's Chief Scientist Jeff Dean and Stanford Professor Fei-Fei Li. The company has already partnered with leading AI-native agent service providers like Decagon, Clay, and Harvey. However, due to the current system's reliance on periodic offline model fine-tuning, achieving real-time online continuous learning on a per-interaction basis still presents a significant engineering challenge.