A multi-agent framework for RePORT India studies.
The LangGraph workflow is orchestrator-driven: a lightweight planning node chooses the next specialized agent based on the current state, recent observations, and any human feedback. This makes the graph flexible and easy to extend with new agents without hard-coding a strict pipeline.
The orchestrator runs in hybrid mode by default: it lets the LLM propose the next action and falls back to deterministic rules when the LLM response is not usable.
- Code Generator: produces Python analysis code.
- Executor: runs generated code on the dataset.
- Error Handler: fixes code issues and retries execution.
- Q&A: answers general questions without code.
- Human-in-the-loop: review checkpoints before and after execution.
- Tool Handler: executes MCP tools requested by agents.
Tool integrations are under active development and may change frequently. The documentation intentionally omits tool-specific setup details until the interface stabilizes.