sia

SIA (Self-Improving AI) is an open-source framework designed to autonomously improve AI performance on benchmark tasks through an iterative loop. It operates by coordinating three main agents: a Meta-Agent that creates a task-specific Target Agent, the Target Agent that executes tasks and records logs, and a Feedback Agent that reviews performance to refine and update the Target Agent. SIA features a live web dashboard to visualize generations, supports multiple LLM providers including Anthropic, OpenAI, and Gemini, and seamlessly integrates with custom datasets like MLE-Bench competitions.

• Self-improving multi-agent loop
• Live web visualization dashboard
• Multi-provider LLM integration
• Automated task execution and evaluation
• MLE-Bench and custom task compatibility