Trajectory, founded by ex-DeepMind researchers, raises $15M seed at $115M valuation
A new platform called Trajectory, built by veterans from Google DeepMind, Apple, OpenAI, and Meta Superintelligence Labs, wants to make AI models learn continuously — not stay frozen after deployment. The company also published a manifesto titled Continual Learning: End of Frozen Software, arguing that AI software shouldn't remain static once shipped. Instead, every user correction, retry, or override should become a direct reward signal that updates the model online.
Trajectory's core mechanism stitches together an agent's execution traces — the precise record of tool calls and intermediate reasoning — with user telemetry that captures whether the final output was accepted, modified, or rolled back. The team points out that these dynamic user interventions have long been thrown away in traditional development, but they are actually gold-standard training signals for judging whether an agent truly solved a complex task.
On the technical side, Trajectory offers a standardized SDK that plugs into popular observability platforms like LangSmith. It automatically imports agent traces and converts conversation history, tool calls, reward feedback, custom metrics, and error logs into a standard Trajectory format for fine-tuning and evaluation. A companion web platform is still under development and will eventually provide visual search and fine-grained management of trace data.
Commercially, Trajectory has closed a $15 million seed round at a $115 million post-money valuation. The round was led by Conviction, with participation from Bessemer Venture Partners, Radical VC, and BoxGroup, along with individual investors like Google Chief Scientist Jeff Dean and Stanford professor Fei-Fei Li. The company has already partnered with leading AI-native agent services including Decagon, Clay, and Harvey. Still, the current system relies on periodic offline model fine-tuning — meaning real per-interaction, real-time continuous learning remains a significant engineering gap.