Multi-Agent Research Lab

Multi-Agent Research Lab is a local agent workflow project focused on state graphs, role collaboration, tool use, memory, verification, and execution traces.

It implements:

• Planner, Researcher, Writer, Verifier, and Summarizer agents
• explicit state transitions
• local tool registry
• lightweight long-term memory
• verifier feedback loop
• reproducible execution traces
• CSV/JSON experiment outputs

Why This Project

Agent projects are strongest when their workflow is observable and testable. This repository shows how a research task can be decomposed into roles, how tools are called, how state is passed across agents, and how verification improves reliability.

Quick Start

PYTHONPATH=src python -m multi_agent_research_lab.cli --question "How should a RAG system improve retrieval quality?"
PYTHONPATH=src python -m multi_agent_research_lab.cli --experiment
PYTHONPATH=src python scripts/run_experiment.py

Current Results

Experiment outputs are stored in:

• experiments/baseline/metrics.csv
• experiments/baseline/traces.json

The baseline reports plan steps, evidence count, tool-call count, and verifier verdict for each research question.

Question	Plan Steps	Evidence Items	Tool Calls	Verdict
RAG retrieval quality	5	4	13	pass
Agent workflow reliability	5	4	13	pass
CTR ranking evaluation	5	4	13	pass

Project Highlights

• Models an Agent workflow as explicit state transitions instead of a hidden loop.
• Separates Planner, Researcher, Writer, Verifier, and Summarizer responsibilities.
• Records tool calls and execution traces for observability.
• Adds a verifier step to make reliability measurable.

Optimization Plan

1. Add branching graph execution and retry policies.
2. Add persistent memory over previous traces.
3. Replace local tools with MCP-style typed tools.
4. Add task success rate and tool-call failure metrics.
5. Rebuild the workflow with LangGraph for framework comparison.

Resume Summary

Built a local multi-agent research workflow with planner/researcher/writer/verifier/summarizer agents, explicit state graph execution, tool-call tracing, memory, and verification metrics.