Overmind: Open-source platform for continuous learning

The platform takes an agent's production traces, scores them as they arrive, and converts them into versioned evaluation and training sets. Teams define what good output means for each capability, and the system benchmarks a small fine-tuned model against the one they already run. Downloadable weights and an OpenAI-compatible endpoint then serve the result alongside frontier models. Self-hosting runs on a single docker compose command, keeping traces and training data inside the buyer's network. The London startup raised a seed round of about £2 million in February 2026 and first pitched itself as a supervision layer for agent security. Observability tools such as Langfuse, LangSmith and Braintrust can feed their traces into it. For an AI company, the case rests on cost and control. A narrow, repeatable agent task can run on a smaller model the team owns instead of a rented frontier API. The AGPL-3.0 license on the platform