buzzbot-ai

Multi-agent AI newsletter pipeline: scouts news, research, tools, video, jobs, and events; runs analysis and prompt engineering; then compiles a Markdown issue with Jinja2. Built with LangGraph and AWS Bedrock (langchain-aws).

What it does

main.py runs a LangGraph workflow with a live Rich status table (per-agent phases and timing). The final Markdown is written under outputs/ as newsletter_issue_{N}_{YYYY-MM-DD}.md, using templates/newsletter.j2.

Pipeline (graph)

Phase	Mode	Agents
1	Parallel from START	news_agent, research_agent, tools_agent, youtube_agent, jobs_agent, events_agent
2	Sequential	analysis_agent → prompt_agent (waits until all Phase 1 nodes finish)
3	Compile	compiler_agent → END (editor’s note, closing thoughts, Jinja render, file save)

Shared state is defined in state.py (NewsletterState): collected sections (news, papers, tools, YouTube, jobs, events, quotes, etc.), LLM-written fields (deep dive, prompt of the week, sentiment), and final_markdown / errors.

Requirements

• Python >=3.13 (see .python-version for the repo default).
• AWS credentials and a Bedrock model ID the account can invoke.
• Tavily API key (web search / news-style gathering in agents).
• YouTube Data API key (video curation).

Setup

1. Clone the repo and create a .env in the project root (same directory as config.py).

2. Set required variables (startup fails fast if any are empty):

   Variable	Purpose
   BEDROCK_MODEL_ID	Bedrock chat model ID
   AWS_ACCESS_KEY_ID	AWS access key
   AWS_SECRET_ACCESS_KEY	AWS secret key
   AWS_DEFAULT_REGION	Region (default in code: us-east-1)
   TAVILY_API_KEY	Tavily
   YOUTUBE_API_KEY	YouTube Data API

3. Optional newsletter metadata:

   Variable	Default
   NEWSLETTER_NAME	THE AI PULSE
   NEWSLETTER_AUTHOR	AI Pulse Team
   NEWSLETTER_ISSUE_NUMBER	1

4. Install dependencies (e.g. with uv — uv.lock is present):

   uv sync

   Or use your preferred tool against pyproject.toml.

Run

python main.py

On Windows, the entry point forces UTF-8 I/O so Rich output (including emoji) is less likely to break the console.

Project layout

• main.py — CLI entry, state init, graph stream, live status
• graph.py — StateGraph wiring
• config.py — .env loading, validation, get_llm() (ChatBedrock)
• state.py — NewsletterState and default_state()
• agents/ — one module per graph node
• tools/ — Tavily, arXiv, YouTube helpers used by collectors
• templates/newsletter.j2 — issue template
• outputs/ — generated .md files

Stack (from pyproject.toml)

LangChain / LangGraph, langchain-aws, Tavily, arXiv, Google API client, boto3, python-dotenv, Rich, Jinja2.