Overmind is an open-source training and optimization platform for continuously improving production AI agents. It scans an agent codebase to map capabilities, prompts, tools, and tasks into a structured graph. OpenTelemetry traces can be converted into versioned evaluation and training datasets tied to those capabilities. The platform runs evaluations, prompt and control-flow experiments, and fine-tuning workflows against real agent behavior. Resulting models can be benchmarked, deployed, rolled back, or accessed through one OpenAI-compatible inference API. Developers can use the web console, CLI, REST API, or MCP integrations for coding agents. It can be hosted by Overmind or self-hosted so traces and training data stay inside the user’s infrastructure.
Features
- Production trace observability
- Versioned training and evaluation datasets
- Prompt and control-flow optimization
- Fine-tuning and model benchmarking
- OpenAI-compatible inference API
- CLI, REST, MCP, and self-hosted workflows