Y Combinator CEO Garry Tan recently dropped a bombshell demonstration video that shook the tech community. As an early core engineer at Palantir and a star founder who sold his startup to Twitter, Garry spent the last two months using Agents to rewrite his former startup project—a project that originally cost $10 million and took 10 engineers two years to build.
More importantly, he open-sourced a framework called GStack and threw out a highly disruptive architectural thesis: "Thin Harness, Fat Skills."
In this article, we will deeply deconstruct the "New Silicon Valley AI R&D Paradigm" shown in the video and see how a human commands a top-tier product development fleet composed entirely of digital lifeforms.
1. The Awakening: AI is No Longer a Tool, But a Regular Army
Garry Tan's sigh at the beginning of the video probably represents the voice of all top hackers right now: "I've coded more in the past two months than I did in all of 2013."
Before this, the industry's perception of AI programming generally stayed at the "Copilot" stage—you write some logic, and AI helps you complete the rest; you encounter a bug, and you paste the error log to AI for analysis.
But in Garry's demonstration, this dialog-box-based "outsourced" collaboration has been completely eliminated.
He astutely pointed out the fatal flaw of current monolithic large models: Because they lack deep contextual memory of your private codebase, if you ask it to directly write a complex system, it will start to "reasonably guess." In the face of a massive codebase, this guessing leads to catastrophic crashes that look perfect but fail on execution.
Since the AI's intelligence is already high enough, why does it still crash?
Garry's answer is deafening: "Humans have never built software by relying on one person. Humans build software through teams, role division, standard operating procedures (SOPs), and code review. Since LLMs are now replacing human work, we must make them follow human team collaboration logic."
This is the background of GStack's birth. It's not a plugin that teaches LLMs how to write code; it's a "Digital Human Organizational Architecture System." It virtualizes product managers, architects, frontend designers, and hardcore backend developers in the terminal, letting LLMs work for you in a team format.
2. Architectural Subversion: Thin Harness, Fat Skills
To achieve this "team-based" collaboration, the traditional approach is to write an extremely massive, rigid Agent framework (Fat Framework). But Garry explicitly points out this is a huge mistake: "LLMs are already smart enough; overly heavy scaffolding will only constrain their potential."
GStack proposes a completely new design philosophy: "Thin Harness, Fat Skills."
- Thin Harness: The underlying scheduler of the system is very lightweight. It only does one thing—maintains the current context in the terminal and hands over the task to the next role at the right time.
- Fat Skills: The real magic lies in the "skill packages" mounted on it. In Garry's demo, each Skill is not a simple API call, but a Domain Specialist with an independent persona and massive internal logic.
This loosely coupled architecture allows you to hot-plug different "digital employees" into the terminal at any time according to your needs.
3. Stage One: The Office Hours Skill (Reshaping the Product's Soul)
In the demo, Garry wants to build a small app to "automatically extract 1099 tax forms from Gmail for users during tax season."
If we followed the old way, we would throw this requirement directly to Claude and let it start writing Gmail API scraping code.
But in GStack, Garry first calls a skill named Office Hours.
This is the most mind-blowing part of the entire video. This skill package encapsulates the soul of 16 top Y Combinator partners coaching founders for tens of thousands of hours. It doesn't start writing code at all; instead, like a picky investor, it continuously throws 6 oppressive "Forcing Questions" at Garry:
- "What is your strongest evidence that anyone actually wants this?"
- "TurboTax already exists, and Plaid can connect directly to banks. Why do you think they need your little tool?"
It doesn't just ask questions; it actively helps the founder iterate the business model.
After multiple rounds of dialogue, this virtual YC partner proposes: "Don't just build a tool to help people download files. We need to use 'finding tax forms' as a Wedge Strategy. The real business model is to funnel the users who have downloaded their tax forms to professional CPAs (CPA Marketplace) and take a cut from it!"
This is the power of "Fat Skills." At this stage, the LLM is not a code generator at all; it is a co-founder with extremely high business Taste. It helps you elevate a tool that can only be sold for $2 into a commercial platform that can charge 10x commission.
4. Stage Two: Adversarial Review
After the business model is finalized, the system enters the phase of writing the Product Requirements Document (PRD).
GStack once again demonstrates the essence of human team collaboration—Code Review and QA.
After generating the initial design document, the system automatically triggers multiple rounds of "Adversarial Review." Another Agent acting as a reviewer starts looking for flaws with a magnifying glass:
- "Your design makes no mention of how to handle the 2FA (Two-Factor Authentication) callback."
- "Missing privacy policy and sensitive data handling statement."
- "Failure handling mechanism is missing."
The most terrifying thing is that after discovering the problems, the two LLMs start to fight each other in the terminal, automatically fixing these 16 pointed-out architectural vulnerabilities. Watching the two code streams flashing alternately in the terminal, a rough idea that originally only scored 6/10 points was forcefully polished into an 8/10 professional-grade technical specification without any human intervention.
5. Stage Three: The LLM Matrix (ADHD CEO and Autistic CTO)
In the visual UI design stage, GStack calls another skill named Design Shotgun.
Here, Garry throws out an extremely vivid and precise metaphor, revealing the true meaning of the LLM matrix:
"You can think of Claude Opus 4.6 as a creative CEO with ADHD. You'd love to grab a beer with him, and his head is full of a billion brilliant ideas and product definitions. But when it's time to actually bite the bullet on the extremely hardcore, boring code implementation, you have to call in the hardcore autistic CTO—and that's OpenAI's Codex model."
This is precisely the essence of GStack's underlying scheduling: No single model can do everything.
When discussing business models and user pain points, call Claude Opus, which is extremely good at empathy and product logic; and when generating specific UI components and underlying hardcore algorithms, the system will instantly and smoothly switch to OpenAI's model.
This practice of perfectly binding the personality traits (Persona) of different LLMs with specific development stages is the ultimate modularity we have always pursued when building the Agent OS.
6. Conclusion: The Age of Discovery for Solo-Founders
After watching Garry Tan's sci-fi-movie-like demonstration, we must admit a fact: The era of the monolithic engineer is over, and the Age of Discovery for Solo-Founders has officially begun.
When you can use a single command to instantly summon a top YC product partner, an extremely demanding QA architect, and a full-stack engineer who works all night without getting tired in the terminal, the marginal cost of software development has infinitely approached zero.
In this cruel dimensional strike, you don't need to write more elegant loop statements than others. The only things you need to possess are the Taste, Vision, and the Audacity to reshape all old rules as the supreme commander.
Top comments (0)