1️⃣ The era of software agents is here—and three new protocols are shaping how they work:
2️⃣ 
MCP, ACP, and A2A.

Think of them as ports, walkie‑talkies, and trade routes for AI. Dive in 👇🧵

3️⃣ MCP (Model Context Protocol)

Open spec from Anthropic.
✅ Standard way to feed prompts, files, APIs to one or many LLMs.
✅ Host ↔ multiple servers pattern.
✅ Great for sandboxed or enterprise‑grade data access.
🔌 “USB‑C for AI.”

4️⃣ When to reach for MCP

• Your LLM needs uniform access to local files, DBs, APIs.
• You care more about context than multi‑agent chat.

5️⃣ ACP (Agent Communication Protocol)

From IBM Research’s BeeAI project.
🪄 Spins up open‑source agents side‑by‑side on your machine.
🗣️ One REST endpoint, deep telemetry, Python + TS SDKs.
⚠️ Pre‑alpha, but promising for local‑first dev workflows.

6️⃣ When ACP shines

• You want a single box running many agents.
• You need rich traces/debug info.
• You don’t (yet) need to mix frameworks outside BeeAI.

7️⃣ A2A (Agent2Agent)

Google‑backed open spec for inter‑framework cooperation.
🗺️ Each agent publishes an agent card (/.well-known/agent.json).
🎫 Tasks flow via simple HTTP; status streams back with SSE.
🔗 Built for LangChain ↔ custom ↔ whatever.

8️⃣ Grab A2A if…

• Your agents live in different ecosystems but must trade tasks.
• You’re building a network of vendor‑agnostic skills.
• You value simple HTTP + discovery over local orchestration.

9️⃣ Can they mix?

🔀 MCP + A2A: feed data to an LLM (MCP) and expose that LLM as an interoperable agent (A2A).
🏠 ACP + MCP: BeeAI could still mount MCP servers for data.
🌐 ACP + A2A: In theory, BeeAI agents could publish A2A endpoints.

🔟 Decision napkin

• Need context ➡️ MCP
• Need local agent farm ➡️ ACP
• Need cross‑vendor collab ➡️ A2A

1️⃣1️⃣ Protocols will keep evolving—pick the one that matches today’s bottleneck, stay modular for tomorrow’s.

1️⃣2️⃣ Deep dive ➡️ “MCP, ACP, A2A, Oh My!” by @zackproser
🔗 https://workos.com/blog/mcp-acp-a2a-oh-my

Software Agents' Protocols

AI Agents That Code Your Game—Meet CrewAI

What is it?

CrewAI is a Python framework for orchestrating multiple AI agents (like a mini dev team) to generate, review, and validate code. Perfect for building something fun—like a Pac-Man or Pong clone—in Python.

Core Concepts

Crews: Groups of specialized agents (e.g., a “Senior Engineer” generating code, QA agent reviewing).

Flows: Event-driven workflows, giving you fine-grained control over who does what, when.

Example Setup

Agents YAML: Defines your game dev, QA, and chief QA roles.

Tasks YAML: Step-by-step instructions (generate code → review → final validation).

Game Design YAML: The actual game requirements (Pac-Man style or whichever idea you fancy).

https://workos.com/blog/how-to-build-a-game-building-agent-system-with-crewai

It’s all run sequentially with a Crew:

1. Code generation by the Senior Engineer agent
2. QA checks by the QA Engineer
3. Final sign-off from the Chief QA Engineer

Get It Running

Create a new Poetry project, install dependencies, and set your OpenAI key.

poetry run python3 main.py will prompt you to choose a game scenario and produce a fully functional Python script.

Why CrewAI Matters

It’s not just for games. 

You can chain specialized agents for anything: blog posts (writer + editor), product specs (engineering + legal), or doc creation (researcher + fact-checker).

Try It Out

Customize your YAML, define your own agent roles, or pick a new game concept—CrewAI will coordinate the entire AI-driven pipeline.

AI Game Coding with CrewAI

The Vercel AI SDK simplifies how you interact with LLMs—no more juggling complex APIs for different providers. 


It’s a single abstraction layer that lets you swap between OpenAI, Anthropic, and more with just a few lines of code.

Latest Release (v4.2): 

Packed with new goodies like:

- Reasoning support

- Model Context Protocol (MCP) integration

- Enhanced handling of complex message parts (think text, sources, reasoning—all in one flow).

Dev Superpower: Instant model swaps. 

Start with GPT-4, switch to Claude or another model—your prompt and code stay almost the same. 

The SDK is backed by infrastructure optimized for AI workloads: 

Efficient streaming, minimal latency, and automatic scaling. 

It’s built for real-world performance, not just proof-of-concepts.

https://zackproser.com/blog/vercel-ai-sdk

Quick Prototyping: Need a streaming text response? Or structured data back from your model? 

The streamText and streamObject functions make it happen. 


UI Made Easy: 

Hooks like useChat handle streaming tokens, conversation history, and more. 

Focus on building an awesome user experience—no need to code up a custom socket solution.

RAG Example: I built a “chat-with-my-data” tool using the AI SDK + Pinecone for retrieval. The pipeline: user question → embed + search → context injection → stream answer + show relevant docs. 

Sure, it’s a new toolkit. 

But once you “get” streaming and how the same interface adapts across multiple LLMs, you’ll be glad you made the investment.

Vercel AI SDK: Model Flexibility

One of my favorite #Nextjs and Vercel features is the Opengraph image support. 

This allows you to *dynamically* assemble an image to be associated with every post that gets "unfurled" when shared on social media and generally on the web by all services that support opengraph

^ This is my updated default OpenGraph image for zackproser.com - so it gets shared whenever someone shares my site: 

I can change the title, description, and even the hero image programmatically for every post: 

This means you can also do things like truncate excessively long titles and descriptions so they don't blow out your image layout.

I wanted to give my card a classic Blueprint feel while still allowing individual posts to show their own flavor and vibe: 


Fun wrinkle: I ship .webp for all my images on my site: 

• ✅ Smaller file sizes: Lossy .webp is ~25–35% smaller than JPEG for similar quality; lossless .webp is ~25% smaller than PNG.
• ✅ Supports transparency
• ✅ Fast — optimized for delivery.
• ✅ Widely supported


However, the OpenGraph image route setup that Next.js and Vercel use doesn't support .webp format right now, so I need to look up the hero image associated with a post and convert it on the fly to .png format before serving the final image response. 

I believe it's worth it!

Dynamic Opengraph Images

🙅 Building a game

☝️😎 Building an #agentic system that builds games on demand for you

https://workos.com/blog/how-to-build-a-game-building-agent-system-with-crewai

How do you get your users to stop re-using insecure passwords that have already been exposed?

Click here.

Reject up to 10 previously used passwords with email + password #authentication.

Read more: workos.com/changelog/password-history http://workos.com/changelog/password-history

⚡🤖 MCP is one of the protocols to watch. 👀 

It makes models like Anthropic's Claude more powerful by connecting them to external services, such as GitHub, Gmail, Research and data APIs, files on your system, etc.

But how do you get started? 

With our free tutorial! ⤵️🧵

We created this step-by-step tutorial for creating a custom MCP server using the official TypeScript SDK. 

You can use the same approach to wire up any API you want so that Claude Desktop can take action on your behalf while you're chatting. 

We cover..

https://workos.com/blog/getting-started-with-claude-desktop-and-custom-mcp-servers-using-the-typescript-sdk

2. Implementing the MCP Server using the official TypeScript SDK

This is an excellent introduction to MCP concepts: 
• Servers
• Tools
• Transports

3. Connecting Claude Desktop

You define MCP servers using JSON in your central Claude configuration JSON - for the ultimate flexibility. 

You can choose how Claude launches each server and with which arguments. 

4. Usage and troubleshooting

Learn what the different Claude Desktop icons mean with regards to registered MCP tools as well as: 

• The MCP Lifecycle - how everything fits together
• Common issues and how to fix them 


This is the same pattern you can use to connect Claude Desktop or Claude Code to any API you want - whether they already have an official MCP plugin!

https://workos.com/blog/getting-started-with-claude-desktop-and-custom-mcp-servers-using-the-typescript-sdk

Harnessing the Power of MCP: A Tutorial

Fine-tuning Llama 3.1:

• Your data
• Lightning AI
• TorchTune

I break down the full process in my latest blog post:
https://zackproser.com/blog/how-to-fine-tune-llama-3-1-on-lightning-ai-with-torchtune

Perfect for anyone ready to dive into custom LLM training.

Launching a SaaS app doesn’t have to be painful.

The Next.js B2B Starter Kit has everything you need to get started:

https://workos.com/blog/nextjs-b2b-starter-kit

LLMs aren’t magic—they’re tools.

If you want them to truly know your data, you need a RAG pipeline.

Here’s the step-by-step guide:
https://zackproser.com/blog/rag-pipeline-tutorial