Autonomous Research Crew (AI Agents)

A multi-agent system that autonomously conducts deep web research, synthesizes findings, and drafts reports—with Human-in-the-Loop (HITL) oversight.

Live Demo

Direct link or use it directly on my portfolio website

The Architecture

This application uses a Cyclic State Graph that schedules 3 AI personas that perform specialized tasks on a common state.

The Crew

• The Researcher: Uses Tavily API to scrape live web data based on the user's prompt.

• The Analyst: Reading the raw data, identifying trends, and structuring the information.

• The Writer: Compiling the analysis into a polished, professional blog post/report.

• The Manager (Human-in-the-Loop): The graph pauses execution before publishing. The user reviews the draft, approves it, or resets the cycle.

Tech Stack & Engineering Decisions

Component	Technology	Thought Process
Orchestration	LangGraph	Enables cyclic workflows and fine-grained state control (unlike standard DAGs).
LLM Engine	Groq (Llama 3.1)	Chosen for ultra-low latency inference (~300 tokens/s) to keep the UI snappy.
Frontend	Streamlit	Rapid prototyping with built-in session state management.
Search Tool	Tavily API	Optimized for LLM agents; returns clean context, not just HTML.
Persistence	MemorySaver	Maintains graph state across Streamlit re-runs (Session Persistence).

Key Features

1. Human-in-the-Loop (HITL)

• The system effectively uses LangGraph's interrupt_before functionality.

  • The AI does the heavy lifting (Research -> Analyze -> Draft).

  • It hits a "breakpoint" and suspends execution.

  • The UI presents the draft to the user.

  • Upon approval, the graph resumes execution from the exact saved state to finalize the output.

2. State Management "The Amnesia Fix"

• Streamlit apps re-run the entire script on every interaction. To prevent the AI from losing its memory or restarting from scratch:

  • I implemented a uuid based Session Threading system.

  • The LangGraph MemorySaver is cached in st.session_state.

  • "Resetting" the agent doesn't delete data; it dynamically generates a new thread_id, instantly giving the user a fresh workspace without server restart.

Local Installation

Prerequisites: Python 3.10+

Clone the repository

git clone [https://github.com/joshgilstrap/research-crew.git](https://github.com/joshgilstrap/research-crew.git)
cd research-crew

Install dependencies

pip install -r requirements.txt

Configure Secrets Create a folder .streamlit and a file secrets.toml:

# .streamlit/secrets.toml
GROQ_API_KEY = "gsk_..."
TAVILY_API_KEY = "tvly-..."

Run the App

streamlit run agent.py

Key Code Snippet: The Graph Definition

# Defining the cyclic workflow
workflow = StateGraph(AgentState)
workflow.add_node("researcher", research_node)
workflow.add_node("analyst", analyst_node)
workflow.add_node("writer", writer_node)
workflow.add_node("reviewer", review_node)

# The logic flow
workflow.add_edge(START, 'researcher')
workflow.add_edge('researcher', 'analyst')
workflow.add_edge('analyst', 'writer')
workflow.add_edge('writer', 'reviewer')
workflow.add_edge('reviewer', END)

🔮 Future Improvements

[ ] Multi-turn Research: Allow the Researcher to self-reflect and search again if data is insufficient.

[ ] Format Selection: Allow users to choose between "Blog Post", "Executive Brief", or "Social Media".

[ ] Export Options: Generate PDF/Markdown file downloads.

🤝 Connect

Built by Josh Gilstrap as a showcase of Agentic Workflows.

LinkedIn

Portfolio