By April 2025, I'd crossed 300+ AI-assisted coding sessions. Deep into Cline, hundreds of weekend projects shipped, and the term "vibe coding" had just been coined by Karpathy. The landscape was shifting fast. I was building an MCP client, experimenting with new frameworks, and watching two entirely fresh protocols emerge in real-time: Model Context Protocol servers (finally gaining traction with HTTP support) and Google's brand-new A2A protocol for agent-to-agent communication.
This episode is where I documented that moment—standing at the intersection of multiple emerging standards, trying to make sense of what frameworks matter and which ones won't.
2025-04-11 · 1:06:38
Episode Highlights
1. From "AI corrects my code" to "I ship without looking" (3:00–5:00)
The journey I've been on these past months surprised even me. In December, I was spending hours fixing and refining AI output. By April, I'd flipped the script completely. I close my eyes and press ship. I let the AI build weekend projects, deploy them live with a handful of users, and validate in production. The economics have changed so radically that perfectionism becomes a liability. You can iterate faster than you can plan.
2. PRD-driven development & test-first workflows (6:50–9:10)
I've been experimenting with a Product Requirement Document generator that takes a simple app description and outputs structured requirements with acceptance criteria. Then I feed that directly into the AI. The dream—still just a dream—is TDD where the AI generates your test cases from requirements, implements the code, and then regenerates the test cases on the next iteration. If all tests pass, your codebase might look completely different, but it's still valid. That's a paradigm shift.
3. MCP servers just got web-friendly (17:00–21:00)
MCP has been around since November, but it hit a wall: it only ran over stdin/stdout, which meant cloning repos locally, running npm builds, and keeping servers running on your machine. In the last few weeks, HTTP support arrived. That changes everything. Suddenly I can host an MCP server on Cloudflare. I can build a "MCP client" (which is really just a chat UI) that discovers and calls remote servers. The deployment story went from hobbyist to production-ready in one update.
4. The middleware question: Are MCP servers just proxies? (20:00–24:00)
Here's what I'm wrestling with: Are MCP servers just thin wrappers over REST APIs? Or are they something more intelligent? I keep thinking about building middleware—smart adapters that sit between the agent and the underlying systems. Servers that understand context, handle rate limiting, maybe do some reasoning about which tool the agent actually needs. The spec is clear on the "what," but the "why" and "how deep" questions are still open.
5. A2A: The new protocol nobody quite understands yet (40:17–44:30)
Google announced the A2A (Agent-to-Agent) protocol alongside a coalition of 50+ companies. It's designed to let agents discover and collaborate with other agents. MCP is tool-centric—"how do I access external systems?"—but A2A is agent-centric—"how do I find and delegate to other agents?" It introduces agent cards (metadata hosted at /.well-known/agent.json), discovery, and long-running task support. Is it competing with MCP or complementary? Google says both can live together, but the real question is: who builds the infrastructure, and does any startup want to be the one-true-A2A-registry?
6. The future of hiring might be the scariest part (13:00–14:30)
My wife—no coding background at all—started vibe-coding with Cline. She shipped features. She got stuck sometimes, needed 30 minutes of help, but mostly moved forward. One year from now, what do we do with the influx of non-developers who can build? The developers who survive won't be the ones who code faster; they'll be the ones who evaluate faster. They can look at output, spot problems, course-correct. That's the superpower.
Where This Connects
This episode sits at a pivot point in the 100-day journey:
- From vibe coding to frameworks — I'd validated that closing-your-eyes-and-shipping actually works. Now: which tools, protocols, and orchestration patterns will stick?
- MCP as infrastructure — Not a toy or a research project anymore. Real deployment options. Real questions about security, discovery, and integration.
- The A2A wild card — Google throwing a new protocol into a space that barely has room for one is either the smartest or most confusing move in AI infrastructure this year.
- Preparing for the May Challenge — Within weeks of this episode, I'd launch into the month-long intensive coding sprint that would define the next chapter.
If you're building with AI, watching MCP, or wondering about A2A, this conversation maps the frontier better than most think pieces will.
This was recorded three months after Episode #76, and the pace of change in those 12 weeks was extraordinary.