🥞 Vercel's New Agent Framework, AI Loops Actually Rock, I'm a Keynote Speaker

Hello Full Stack PMs!

Welcome back to Full Stack Friday, serving up the hottest AI developments fresh off the griddle for product people. 🥞

I really didn’t manage my time well and finished this up at the USA v Australia World Cup match (which is happening in Seattle, where I am!), so we’ll keep this snappy. But there’s still some great stuff I wanted to share.

Yes, Fable was banned. There’s not a lot more to say about that, so we won’t dwell on it. While we wait for its return, we’ve got a whole new agent framework from Vercel and a whole new agent meta to explore with loops.

Here’s the plan:

Yes, I wish I’d finished this yesterday 🙂

Let’s do this.

🎓 My Claude Code Mastery course is open and full of great people. Claude Code for PMs: Mastery has 750+ product folks in it, an active Slack, and two weekly office hours. It’s a membership, so one price gets you every course I make, and Cursor and Codex tracks are already on the way. If you’ve been waiting for a sign to jump in, this is it.

🇳🇱 I’ll be the keynote speaker at ProductCamp Amsterdam! It’s on Saturday, Aug 29. My topic is Agentic Product Management, but it’s otherwise a “product unconference” where the agenda gets pitched and voted on by the room the morning of. It’s going to be a lot of fun. If you’re in the Netherlands or nearby, come through!

🔗➡️ Early bird tickets are still available for €75

🏗️ The next Team OS workshop is July 29 with Hannah Stulberg (PM building this for real @ DoorDash), and it’s the most hands-on thing I run: real working demos for every concept you learn along the way. Enroll now and you also get the recording of the cohorts that already ran plus every future cohort (you always get access to all future sessions!), so today is the most you’ll ever get for the price.

🔗➡️ Grab a seat, $250 for July 29, 5:30–8:30pm EDT

eve is Vercel’s new open-source “Next.js for agents,” a framework for building, running, and scaling production agents that launched June 17 at Vercel Ship. The mental model is that an agent is a directory of files: instructions in markdown, tools, skills, channels, and schedules in conventional folders, all auto-discovered. A minimal agent is about 2 to 3 files.

What I love about it is how directly it maps to working with agentic tools like Claude Code or Cursor, where claude.md, skills, rules, and hooks are “all just files.” If you’re already familiar with those tools, eve is straightforward: you define what it can do, how it does it, and the instructions for doing it. And it hands you everything you need to make these agents production ready, like durable execution, sandboxing, and human-in-the-loop approvals that pause indefinitely without burning compute.

Since it’s Vercel, it’s all built FOR the agents you already use to help you build it. If you’ve done my Vibe Coding module, you know how easy deployment gets when you give your AI access to the Vercel CLI. It’s still in beta, but I’ll be sharing some detailed guides on it soon.

Good resources:

A new meta is evolving around setting up your AI agents to run in continuously improving loops. Back to back, Boris Cherny (Claude Code lead) said he just writes loops for his agents, and Peter Steinberger (creator of OpenClaw) put it like this:

The same idea is showing up at the model level. People are getting strong results from a “Fusion” of models, fanning a single prompt out to several at once and synthesizing the best answer. In OpenRouter’s own benchmark, pairing Opus 4.8 with GPT 5.5 actually edged out a solo frontier model:

I was skeptical at first (it sounds like a great way for these companies to get you to spend a lot of tokens!), but after testing it with Opus 4.8 on ultracode paired with the Codex plugin using the /adversarial-review skill, I am a believer! In my experience, Opus 4.8 tends to understand intent better and be more creative, but that also means more hallucinations and mistakes. GPT 5.5 is a better executor, sometimes to a fault. By forcing Opus to have its work checked by GPT 5.5, I almost feel the same kind of trust I had with Fable.

For your next ambitious project, try setting up a review loop with clear exit criteria and see if you can step out of the loop more than usual. I’m running the $100/mo Claude Max and $20/mo ChatGPT plans.

Good resources:

If you had a chance to use Fable, you’ll relate to this:

Whatever wins, your connection to all of it is going to be through a coding agent.

New here? Start with the free courses: Claude Code for PMs, Cursor for PMs, and Claude Code for Everyone.

Done those? Come join Mastery for the full curriculum, the Slack, and weekly office hours. Still $200 with EARLYAPPLICANT.

And if you lead a team, the next Team OS workshop with Hannah is July 29. Join now and you’ll get access to the past cohort right away.

Keep building,

Carl 🥞