Hey {{first_name | there}},

You sent one message. Twelve AI agents read it, debated it, delegated it, verified it, and sent back one response. Not just one, but twelve.

That's Fugu.

Sakana AI just publicly launched Fugu, and it’s already competing frontier models like Fable and Mythos on reasoning, science and engineering benchmarks.

Here's what Fugu actually is: a single API endpoint that, under the hood, runs an entire coordinated team of AI models. It picks the best model for each sub-task, delegates, verifies results, combines them into one final answer. 

But here's the wild part- it can call recursive copies of itself 

But the pricing isn't the interesting part. The architecture is.

For the past two years, anyone building AI workflows has been doing the same thing: manually scripting roles, sequencing models, wiring agents together with code. 

You decide which model does what. You define the flow. You babysit the pipeline.

Fugu flips that entirely. It trained a model to decide all of that dynamically, who to invoke, what role to assign, how to split sub-tasks, for each specific request. There's no static workflow. The orchestration is learned, adaptive, and invisible.

This is a significant architectural shift. The bottleneck was always the human in the middle deciding how to wire models together. It’s gone now.

And that brings me to something Dario Amodei said that I haven't been able to stop thinking about.

Amodei, the CEO of Anthropic, said it plainly: coding is the first profession AI will replace. Not eventually. The broader task of software engineering takes a little longer. 

But the entry point, the part where a human manually scripts how logic flows, is already going.

Fugu is a live demonstration of why he's right.

But here's where it gets more unsettling and more interesting.

Back then Anthropic ran internal studies and found something most people glossed over: depending on how you use AI, measurable de-skilling in coding is real and observable.

The tool doesn't cause it. Carelessness does.

This is the paradox I keep running into: AI makes you 20x more productive at tasks, but only if you understand the 5% you're still responsible for.

If you outsource that 5% too, you don't become more productive. You become replaceable.

The research even found a ~14% drop in hiring for workers aged 22–25 in AI-exposed occupations. That's not the future. That's now.

So what does this mean practically?

If you're building AI products right now, agents, workflows, automations, here's the uncomfortable question:

Are you building pipelines that a trained orchestration model will make obsolete in 12 months?

Because Fugu's architecture isn't a product feature. It's a direction. The direction is: humans shouldn't be manually deciding model orchestration. Systems should learn to do that themselves.

This doesn't mean stop building. It means build at the layer Fugu can't reach the layer of judgment, taste, context, and human-shaped problems that no amount of recursive self-delegation resolves.

Amodei's actual advice for what survives: human-centered professions that mix people, the physical world, and analytical thinking. The radiologist who understands patients, not just reads scans. The builder who understands what the user actually needs, not just what they typed.

Critical thinking, he said, might be the most important skill of the next decade. When AI can generate anything, the ability to tell what's real from what's not becomes rare and genuinely valuable.

For two years, we've been thinking about AI as: one model, one brain, one response. Fugu quietly obsoletes that mental model. The future looks more like: one request, invisible company, one answer.

The question isn't whether that's coming. It's already here, at $5 per million tokens.

The question is: what role do you play in that company that you can't be automated out of?

I'd genuinely love to know what you think. Reply and tell me, are you building at the orchestration layer, or the judgment layer?

— Aashish

Every Wednesday at 5 PM CT, Gauntlet AI professors teach a live, hands-on AI engineering session — completely free. If you're nontechnical, this isn't for you. New topic every week, built for engineers who want to build, not just watch. See upcoming sessions.

Browse Free Sessions