Research agents that think in schemas, not stream-of-consciousness

What it does

SGR Agent Core is a Python framework for building research agents around Schema-Guided Reasoning. It offers an extensible BaseAgent interface built on a two-phase architecture, plus concrete implementations like SGRAgent and SGRToolCallingAgent. The library bundles search, reasoning, and clarification tools into an OpenAI-compatible REST API that streams responses via SSE and works with local or remote LLMs.

The interesting bit

The framework attempts to split agentic work into structured reasoning and flexible tool selection, asking the model to think in schemas before it acts. That two-phase split is meant to reduce the meandering common in pure tool-calling or ReAct loops by enforcing intermediate structure without locking the agent into a rigid sequence.

Key highlights

• Three built-in agent types: SGRAgent, ToolCallingAgent, and SGRToolCallingAgent.
• OpenAI-compatible REST API with SSE streaming; works with local models for fully private research.
• Interactive CLI (sgrsh) handles single-shot queries, agent selection, and chat mode.
• Benchmarked on SimpleQA at 86.08% accuracy with gpt-4.1-mini.
• Extensible tool interface for search, reasoning, and clarification.

Verdict

Worth a look if you are building research pipelines and want a self-hosted, structured alternative to standard tool-calling agents. Skip it if you need a managed end-to-end service or a no-code interface—this is a library that expects you to write config files and custom tools.