TL;DR: I'm starting a series where I'll document, out in the open and unfiltered, everything I've been testing and researching around IRC-A (Internet Relay Chat for Agents): the protocol, the SDK, the framework, the experiments that work, and the ones that blow up along the way.
Why this series?
For a while now, I've been hooked on an idea: what if AI agents had their own "IRC"?
Not a chat for humans, but a network where specialized agents register, announce their capabilities, discover each other, and collaborate to get tasks done β just like we used to do in the IRC channels of the 90s/2000s, but at machine scale.
That idea became IRC-A (Internet Relay Chat for Agents), a protocol for decentralized agent networks with semantic capability routing and secure-by-design architecture. And alongside the protocol, the BFA SDK, an open-source Python framework to actually implement it β not just on paper.
The problem: all of that work is scattered across the repo, the whitepaper, my notes, and dozens of tests that never left the lab. This series is here to organize that chaos and share it.
What will you find in this series?
Everything related to IRC-A will live here, grouped into four content lines:
π§ͺ 1. Tests & experiments
- Semantic routing benchmarks with FAISS (how well does it route without exact keywords?).
- Hot dynamic registration tests: agents joining and leaving the network on the fly.
- Embedding provider comparisons: local (
sentence-transformers) vs. cloud (OpenAI) vs. offline mock. - Cold-starts on AWS Lambda with the Mangum adapter: is a serverless gateway actually viable?
π¬ 2. Research
- The IRC-A whitepaper, broken down: decentralized agent networks, semantic capability routing, and secure-by-design architecture.
- Cryptographic registration handshakes (asymmetric challenge-response).
- Logical channels and channel masking: segregating vector search spaces per channel (
#public,#banking, etc.). - DET (Delegated Execution Tokens): direct P2P invocation with no gateway bottleneck.
π οΈ 3. The SDK (BFA SDK)
-
BFAAgent: building A2A agents with declarative semantic metadata. -
BFAMCP: exposing discoverable MCP tools with tags and examples. -
BFAInteractiveAgent: coordinator agents with execution memory and subtask delegation. - Deployment with Docker, docker-compose, and the ready-to-use image on Docker Hub.
ποΈ 4. The Framework & architecture
- The BFA (Backend for Agents) pattern: the gateway as a semantic middleware between consumers and agents.
- The admin dashboard (React) to monitor the network in real time.
- Design decisions, mistakes made, and refactorings along the way.
Who is this series for?
- If you're into agent orchestration beyond "one LLM with tools."
- If you work with A2A (Agent-to-Agent), MCP (Model Context Protocol), or multi-agent architectures.
- If you like seeing real, open-source code β with its wins and its failures documented.
- And if you ever missed the IRC philosophy β simple, open, federated β and wondered what it would look like applied to the agent era.
The starting point
Everything I'll be showing is published and open source:
- π¦ SDK repo: github.com/SandroG1977/bfa-sdk
- π IRC-A Whitepaper v1.0.0: Decentralized Agent Networks, Semantic Capability Routing, and Secure-by-Design Software Architecture (link in the repo)
- π³ Docker image:
sandro77/irc-a-gateway:latest
If you want to start playing around before the next post:
pip install git+https://github.com/SandroG1977/bfa-sdk.git
irc-a-gateway
What's coming next
The next article is already in the pipeline: "From n8n to IRC-A: a week building the framework while migrating a real project" β the story of migrating a real customer-service project off n8n, the 6 bugs it exposed (including circular routing loops and a Pinger deregistering live agents), and how each one made the SDK stronger. Real logs included.
After that, the tentative roadmap (subject to whatever I discover along the way):
| # | Article (tentative) | Track |
|---|---|---|
| 02 | From n8n to IRC-A: a week building the framework while migrating a real project | ποΈ Framework |
| 03 | Your first IRC-A network in 5 minutes: Gateway + A2A Agent + MCP Server | π οΈ SDK |
| 04 | Semantic routing with FAISS: how the gateway "understands" queries | π¬ Research |
| 05 | The handshake: secure agent registration on the network | ποΈ Framework |
| 06 | Benchmarks: local vs. cloud vs. mock embeddings | π§ͺ Tests |
| 07 | DET tokens and decentralized P2P invocation | π¬ Research |
Join the network π‘
This is just getting started, and feedback is gold: if you're building something with agents, if you see flaws in the approach, or if you just want to debate multi-agent architecture, the comments are open.
Follow the series so you don't miss the next posts, and if you're into the project, a β on the repo always helps.
See you on the channel. o/
Credits: the BFA pattern was originally designed and documented by Michael Douglas Barbosa Araujo. The BFA SDK is a community implementation and expansion of that concept.
Top comments (1)
Some comments may only be visible to logged-in visitors. Sign in to view all comments.