Newsletter multi-agents
From monitoring to newsletter, orchestrated by AI agents.
A multi-agent system that scans AI news sources (vendor blogs, arXiv, Reddit, Substack, web search), groups articles by sub-topic, spots trends and writes a newsletter ready for review. An internal BotiqueAI project: the engine is real code, the demo interface uses sample data.
Keeping up with AI news
without losing your days.
AI news is scattered across blogs, research papers and social networks. Following it by hand takes hours and the result is a pile of links rather than a synthesis.
A graph of agents
run by a supervisor.
An LLM supervisor interprets the request (topic and period) and routes it to the search agent, which calls the collection tools. Results are clustered, summarised, turned into a post and then analysed for trends.
The supervisor extracts the topic and date from the request using structured output and picks the next step: search, answer a question or end the session. The search agent then calls the collection tools.
Articles from AI vendor blogs, arXiv, DuckDuckGo, Reddit and Substack, with page content extraction and LLM relevance filtering.
Articles are clustered by sub-topic, each cluster summarised, then a post is written with links and cited sources.
A trend-analysis node completes the post, and the output (newsletter, clusters, articles) is saved as JSON. A daily cron-scheduled script re-runs collection and generation.
A monitoring chain
that runs by itself.
Python, LangGraph
and a light LLM
A LangGraph state graph, a gpt-4o-mini model driven by LangChain, source connectors and two interfaces (API and web app).