Hefesto
AI team orchestration
Hefesto orchestrates teams of persistent AI agents that move tasks forward autonomously across a Kanban board (ToDo → InProgress → InReview → Done). Each task passes through QA and Security reviews before being merged for real into main, mirroring the workflow of a human development team. The system combines an agentic engine built on Semantic Kernel, Ollama models running locally or in the cloud, and retrieval-augmented generation (RAG) over Qdrant, all executed inside an isolated sandbox on Kubernetes.
What it does
Hefesto turns a Kanban board into a team of autonomous agents that pick up tasks, build them, review them and merge them without constant manual intervention.
- Orchestrates teams of persistent AI agents that advance tasks autonomously across the ToDo, InProgress, InReview and Done columns.
- Runs automated QA and Security reviews before any change is approved.
- Performs real code integration into the
mainbranch once the reviews pass. - Retrieves relevant project context through RAG over a Qdrant vector store.
- Lets you switch between local or cloud Ollama models depending on the workload.
- Isolates agent execution in a Kubernetes sandbox to keep runs secure.
Architecture
The platform is structured as 15 .NET 10 projects with an event-driven design that decouples the agentic engine, the board orchestration and the review services.
- 15 .NET 10 projects organized with an event-driven architecture.
- Agentic engine implemented on Semantic Kernel to coordinate agents and their tools.
- Language models served by Ollama, interchangeable between local and cloud execution.
- RAG backed by a Qdrant vector database for project context.
- Isolated Kubernetes sandbox to run agent work safely.
- A 6-stage GitLab CI/CD pipeline for build, testing and integration.
Stack
Built with a 15-project, event-driven .NET 10 backend in C#, a Semantic Kernel engine, Ollama models and RAG over Qdrant, a Next.js 16 frontend in TypeScript, Kubernetes execution and a 6-stage GitLab CI/CD pipeline.