SELECTOOLS: Multi-agent graphs, tool calling, RAG, 50 evaluators, PII redaction
Selectools is an open-source Python framework for AI agent systems, supporting OpenAI, Anthropic, Gemini, and Ollama. It introduces new features like multi-agent orchestration, a built-in evaluation framework, and autonomous bug hunting.
Why it matters
Selectools provides a robust, open-source framework for building and evaluating multi-agent AI systems, with advanced features like interrupt handling, orchestration, and automated testing.
Key Points
- 1Selectools uses Python generators for interrupt handling, avoiding the need for idempotent side effects
- 2AgentGraph provides deterministic control flow with LLMs handling reasoning within nodes
- 3Supervisor Agent supports 4 coordination strategies for structured tasks, iterative refinement, heterogeneous tasks, and autonomous research
- 450 built-in evaluators (30 deterministic, 20 LLM-based) with A/B testing, regression detection, and reporting
Details
Selectools is an open-source Python framework for building AI agent systems. It introduces several key features: 1) Interrupt handling using Python generators, which resume at the exact yield point instead of restarting the entire node. This avoids the need for idempotent side effects. 2) AgentGraph, a directed graph executor for agent nodes, with deterministic routing and control flow. This prevents context explosion and supports parallel execution. 3) Supervisor Agent with 4 coordination strategies for different task types. 4) A built-in evaluation framework with 50 evaluators (30 deterministic, 20 LLM-based) for A/B testing, regression detection, and reporting. The project also includes autonomous bug hunting, with a convergence system that runs 8 passes across all modules until 3 consecutive clean passes, resulting in 90 bugs fixed and 254 new regression tests added.
No comments yet
Be the first to comment