Rafael COELHO
← Work

FEBRUARY 2025

COELHO Agents

Multi-agent platform combining a memory-aware assistant, an autonomous Software Developer agent with self-fixing code loops, and a GraphRAG-based YouTube Content Search — all orchestrated with LangChain + LangGraph.

Outcome
Three production agent tools shipped end-to-end · autonomous code-fix retry loops in the Software Developer · GraphRAG (Neo4j) backing multi-hop video query reasoning.
LangChain LangGraph Neo4j GraphRAG Streamlit Python LLM Apps Multi-agent systems

Executive summary

COELHO Agents packages three production-grade agentic tools behind a single Streamlit interface, all orchestrated with LangChain and LangGraph:

  1. Simple Assistant — chatbot with memory and real-time response
  2. Software Developer — autonomous coding agent that generates code, runs it, and fixes its own bugs in retry loops
  3. YouTube Content Search — GraphRAG-based video query system backed by a Neo4j knowledge graph (the research that became COELHO Nexus)

The project demonstrates three distinct agentic patterns — conversational memory, plan-execute-diagnose-patch loops, and graph-augmented retrieval — running on the same LangChain / LangGraph foundation.

See it in action

The agents demand a working Python environment, multiple LLM API keys, and a running Neo4j instance, so you can’t just spin them up. This walkthrough is the verifiable record of the three tools in operation — memory-aware chat, autonomous code-fix loops, and live GraphRAG video queries.

Open on YouTube ↗

Three agentic tools, three distinct patterns

ToolAgent patternWhy it’s non-trivial
Simple AssistantConversational memory over LangChain message historyPersistent context across turns — not stateless one-shot prompts
Software DeveloperPlan → write → execute → diagnose → patch retry loop (LangGraph state machine)Self-correcting code generation: the agent runs its own code, parses errors, rewrites, retries — autonomous bug fixing
YouTube Content SearchGraphRAG over Neo4j knowledge graph extracted from video transcriptsMulti-hop reasoning impossible with vector RAG alone

Why Knowledge Graph + Agents > vanilla RAG

Vector RAG retrieves by semantic similarity — fine for paraphrase-level questions, but it can’t answer multi-hop relational queries like “who has been interviewed about LLM safety on this channel?” Knowledge Graphs encode entities and relationships explicitly (X works_for Y → Y located_in Z), so a Cypher query walks the graph and the LLM grounds its answer in real relationships rather than chunk-similarity lottery.

The YouTube Content Search tool here was the first prototype of this pattern. The operational primitives — agent loops, graph extraction, multi-hop traversal — became the architectural backbone of COELHO Nexus.

Stack

  • Streamlit — interactive interface
  • LangChain — agent and tool composition primitives
  • LangGraph — multi-agent orchestration with explicit state graphs and retry edges (powers the Software Developer’s self-fix loop)
  • Neo4j — Knowledge Graph storage + Cypher querying

What this project proves

  • Three distinct agent patterns shipped in one codebase — memory-aware conversation, autonomous code-fix loops, and graph-augmented retrieval — not toy demos
  • LangGraph state machines work for real production agent loops — including self-correcting code generation with retry semantics
  • The GraphRAG pattern scales — what started here as YouTube transcript search became the production retrieval backbone of COELHO Nexus

Source on GitHub →