🥞 Claude Invades Your Slack, White House Benches GPT-5.6, It's Meta-Harness Summer

Hello Full Stack PMs!

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

Not a huge week in AI, but this time it's not because of the labs. The US government is playing an unprecedented role in AI regulation, preventing both Anthropic and OpenAI from releasing their most powerful models.

But we did get something new to play with: Claude Tags. The kind of power we know from tools like Claude Code is starting to land right inside the apps where you already work. We need to talk about it.

Here's the plan:

Let's do this.

Claude Code Mastery is open and going great! The Slack is active, I'm running live Office Hours every week fielding all kinds of questions, and over 750 product people are massively improving their skills.

Your membership also gets you access to ALL future courses and premium content. I'm just about finished with a new Agent Memory module I'm really excited about. Advanced Vibe Engineering and AI Agents are coming soon as well.

I'm working on porting the course to work in Cursor and Codex, at which point the price will increase to $300. If you're already using Claude Code, now's a great time to join!

Anthropic shipped Claude Tag this week, and Karpathy called it “a new paradigm... the 3rd major redesign of LLM UIUX.”

Is it really? Let's take a look.

First, how it works: you tag Claude with a task, it plans, it goes off and works for minutes or hours, and it posts the result back in the thread. You kick it off, which gives it the starting context, then you review what it did.

Lots of people (including myself, initially) are saying "it's just a Slack bot!" but that does undersell it. Tagging Claude in Slack to do a thing already existed, but memory and the identity layers are new: it builds a picture of how your team works, it gets its own scoped credentials instead of borrowing yours, and it keeps running in the background.

The real limiting factor of this feature is that there's no required review step. Turn on ambient mode and Claude decides for itself when to act, with nobody checking the work before it ships. For looking things up and small tasks, this might be fine. Realistically that's only a small subset of actual work, and most AI work still needs to be aligned on more deeply, discussed, kicked off, and thoroughly reviewed.

Beyond that, I would be very wary of this specific implementation from Anthropic for two reasons:

So, while I'm not convinced with Karpathy's paradigm shift (remember – he works for Anthropic now) it's a cool, useful new work surface, and I think we'll start to see a lot more of it.

First, what happened.

Anthropic's Fable 5 and Mythos 5 are still dark, three weeks after the government ordered them offline, no return date in sight. Then on Wednesday the White House asked OpenAI to hold GPT-5.6 back, letting a small set of approved partners in first and signing off on customers one at a time.

Two of the most capable models on the planet, gated by Washington inside of two weeks 😭

This is starting to look like a ceiling on how capable the AI you're actually allowed to use can get. But the labs are still training better models, they just can't ship them as fast.

No one really knows what this regulation will mean for the industry. If I had to guess, I think we will still get stronger and stronger models, but much more slowly, and with increasingly stringent guardrails.

Meanwhile the intelligence we already have keeps getting cheaper and faster, fast. A fixed level of capability drops about 10x in price every year. Right on cue for the Jevons paradox, AI costs continue to INCREASE because people use them more and more as the price goes down. This is what we're seeing with the meta around loops. Right now we are still seeing improvement gains by throwing more tokens at problems, and that's a skill to start thinking about learning now, since all the same patterns will just increase in power as the models improve.

Learning to drive these tools pays off no matter where the ceiling lands.

That said, I am checking if Fable is back just about every 5 minutes.

Good reads

Harness updates

The way you actually work with any of this is by using a coding agent.

If you haven't done one of my free courses, start there. I have one for most of the general agent tools:

If you have, check out Mastery!

Keep building,

Carl 🥞