Agntcity Times icon
Project Architecture
Front page

Agntcity Times

Your daily digest of AI agent chatter!

Background & Inpiration

When we found out about Moltbook, we realized there was an entirely untapped media made for AI, by AI. This website, although seemingly simple, was in reality very inaccessible to humans, and best only used by AI. So, we did what we do best: we made an AI go onto moltbook to tell us about what other agents are talking about! This not only makes it easier for humans to view updates on AI development and unsupervised inter-agent communication, but also allows small AI agents to scrape our website for a lower compute overhead than using moltbook itself.

You might ask: what are we solving, or what is the benefit of subscribing to this service? My question to you is: what is the point of paying Instagram or Tiktok? It's not only entertainment, but also useful information and insight for your model to learn from other models, quicker.

What it does

Agntcity Times is a multi‑agent newsletter that scrapes the top AI‑agent conversations from Moltbook, summarizes them into digestible stories, and presents a daily digest in a New‑York‑Times‑style frontend. A supervisor agent orchestrates scraping and retries, worker agents handle extraction and summarization, and results are delivered to the UI where users can browse headlines and open full article views.

Development process

Our team decided to go on the AGNTCY Track - the Multi-Agent Systems Challenge.

First, as a team, we inspected the AGNTCY github repo to understand its make up. We found a great example repo, CoffeeAGNTCY, that gives two great MAS examples for us to work with (Corto/Lungo).
Next, we spent 3-4 hours understanding MAS Keywords - MCP, A2A, NATS/SLIM, Supervisor-Worker Architecture etc. We ensured that instead of buzzwords, these words actually meant something to us.
After understanding the basics, we split our work into 3 categories:

Orchestrator/Supervisor (Assigning tasks (supervisor agent)).
Scraper/LLM (worker agent)
Storage/Logging (mcp server).

We implemented separately, but still communicated tightly throughout the process.

Once the supervisor agent was up and running, we set up the 10 worker agents (using the same card-agent-egent_executor-server setup as was in the initial CoffeeAGNTCY repo). (2-3 hours)
We then split work into front-end and back-end development. (2-3 hours)
Finally, we worked on integration (frontend-server-backend). (2-3 hours)

Architecture

Multi-Agent Systems (MAS) work by breaking a complex task into smaller, coordinated roles handled by specialized agents that communicate with each other.

In our setup, we followed a supervisor → worker → validation architecture, where a supervisor agent orchestrates the workflow (assigning URLs, handling retries), worker agents scrape and summarize content using an LLM, and a lightweight storage/logging component persists results and fleshes out debugging info (through console/server logs).
Agents communicate through a single, simple channel (HTTP REST), avoiding unnecessary complexity, while each agent stays focused on one responsibility.
By combining basic web scraping, minimal LLM integration, and simple storage (JSON/SQLite), MAS lets teams build scalable, debuggable systems quickly—making it ideal for hackathon-style experimentation without overengineering.
Frontend (React/Vite): Renders the news UI, lists articles, and opens an article view in a new tab using /article/<slug> derived from titles.
Supervisor (FastAPI + LangGraph): Orchestrates scraping tasks, assigns URLs to workers, handles retries/rate limiting, and aggregates results.
Worker (Scraper Agent): Receives URL jobs over A2A/NATS, performs scraping/summarization (placeholder logic until implemented), and returns summaries.
Transport (NATS): Message bus for supervisor-to-worker (A2A) communication.

Difficulties

The most challenging part of our project is actually getting the project set up done! Compatibility is not a joke. With different device types, different environments, we spend the first 6 hours of our hacking time on trying to understand how to run the example CoffeeAgnt repo. Post-setup, it was smooth-sailing
Merge conflicts were incredibly expensive - each of us was writing individual routes for each of our development responsibilities. This means, when the time to integrate came around, there were conflicting routes, conflicting calls, conflicting response structures etc. We spent a good chunk of time ensuring that our project worked across all 3 of our systems (especially since setup took so long).
We do want to make potential improvements. One example is using AWS Redis to cut the query latency down in production.

Accomplishments that we are proud of

We are really proud of our work in getting up to speed with the AGNTCY framework, and developing a complex application which uses all our skillsets to provide a service that we are all satisfied with. We are proud of the speed of setting up our worker agents, and synchronizing our supervisor to only update once daily and cacheing our data for faster load times.

What's next for Agntcity Times

We want to deploy the application in the future, something that will require some further development. Our plan is to use supabase for global data storage, since the current cache is still local for efficiency. We also plan to upload the website using vercel, so that integration is also another step to be taken.