1. Headline & intro
YC boss Garry Tan hasn’t just open-sourced some prompts – he’s turned his personal AI coding setup into a minor cultural event. His "gstack" configuration for Claude Code has triggered awe, ridicule, and yet another round of existential questioning about what software work looks like in the agent era. You don’t need to care about Tan personally to care about this moment: it’s a live demo of how power users, platforms and hype collide. In this piece we’ll look past the drama and ask what gstack really tells us about AI agents, developer work, and who wins when a celebrity founder open-sources their brain.
2. The news in brief
According to TechCrunch, Y Combinator CEO Garry Tan used a SXSW stage to describe how obsessed he is with AI agents, saying he sleeps only a few hours because he’s so energized by working with them. Two days before that talk, he published his personal Claude Code configuration – called "gstack" – on GitHub under an open-source license.
Gstack is essentially a set of opinionated "skills": reusable prompt workflows stored in markdown files that tell Anthropic’s Claude Code how to act in specific roles (CEO, engineer, reviewer, designer, documenter, and so on). The repository quickly went viral, amassing almost 20,000 stars and more than 2,000 forks, and trending on X and Product Hunt.
The backlash came when Tan claimed, again on X, that a CTO friend used gstack to quickly uncover a security flaw in production code. Critics argued the story was either exaggerated or reflected badly on that team’s security practices. TechCrunch notes that even other AI models like Claude, ChatGPT and Gemini evaluated gstack as a solid, sophisticated workflow – but not magic.
3. Why this matters
Gstack is not groundbreaking technology. It’s a configuration of prompts and roles that many heavy AI users have independently discovered. The reason it matters is who released it, and what it reveals about the emerging "agent-native" way of working.
The winners in the short term are:
- Anthropic and the AI tooling ecosystem. Each high-profile workflow like this is a free product-marketing campaign for Claude Code and the idea of agent teams as the new IDE.
- Non-expert founders and solo builders. Gstack distills a pattern that seasoned AI tinkerers already use: don’t ask the model for "the app"; simulate a small company with specialized roles that critique each other. That’s real leverage for people who can’t yet orchestrate this themselves.
Who loses?
- Traditional team boundaries. When a celebrity founder publicly describes ten AI "workers" shipping features, it normalizes the idea that a lot of junior-to-mid-level work is at risk of being automated or compressed.
- Developers who are already skeptical. To anyone writing code daily, a viral set of markdown prompts from a well-connected investor feels like yet another example of hype drowning out quieter, more nuanced progress.
Most importantly, gstack surfaces a deeper shift: AI productivity is moving from "ask-and-reply" chats to process design. The real skill is not knowing Python or React, but knowing how to structure a workflow where multiple agents with different instructions collaborate, check and improve each other. That’s closer to management science than to classic coding.
4. The bigger picture
Tan’s setup sits at the intersection of several trends we’ve been watching since 2023:
From copilots to teams of agents. GitHub Copilot and similar tools started as autocomplete-on-steroids. By 2024, we saw the rise of "AI engineers" and agent frameworks promising to own entire tickets. Gstack is a user-land version of the same idea: instead of a single assistant, you orchestrate a mini-org chart of specialists.
Configuration as a new kind of IP. In the cloud era, companies guarded their infrastructure as code. In the AI era, the secret sauce is increasingly workflows – the combination of prompts, tools, and review loops. Tan open-sourcing his stack is a bit like publishing your company’s internal playbook. It raises the question: what remains proprietary when everyone can copy your best patterns overnight?
Celebrity amplification in open source. Thousands of developers have posted similar prompt stacks that went largely unnoticed. When the head of Y Combinator does it, it rockets to the top of GitHub and Product Hunt in days. That doesn’t make gstack useless – but it does remind us that visibility in the AI tooling world is heavily skewed by social capital.
Compared with other ecosystems – from VS Code "devcontainers" to Cursor-style AI-native IDEs – gstack is relatively lightweight. It’s more cultural artifact than engineering feat. Yet that’s exactly why it’s powerful: it makes the idea of a virtual team of agents tangible enough that non-technical CEOs suddenly feel they "get" how to use AI for real work, not just for chat experiments.
5. The European / regional angle
For European readers, two aspects stand out.
First, culture and founder health. Tan’s description of sleeping four hours a night because AI agents keep him wired feels straight out of classic Silicon Valley hustle mythology. In many European ecosystems – from Berlin to Ljubljana – investors and founders have spent the last years trying to move away from glorifying burnout. When the most visible accelerator boss on the planet celebrates AI-fuelled insomnia, it risks dragging the conversation backwards.
Second, regulation and data protection. Claude Code and similar tools are usually hosted outside the EU. Any European startup copying gstack to work on real customer code must immediately think about:
- Where does the code go? Does it contain personal data covered by GDPR?
- How does this interact with the upcoming EU AI Act, which will likely require transparency and risk management for systems that meaningfully impact users?
There is also a strategic angle: Europe historically lags in foundational models, but is competitive in developer tools and enterprise software. A workflow-centric approach plays to European strengths: rigorous processes, compliance-aware engineering, and industry-specific solutions. We’re already seeing early European players packaging AI agents into regulated-friendly offerings, which may feel less flashy than gstack but more aligned with how EU companies actually buy software.
Finally, the gstack debate is a reminder that European open-source work often gets less global visibility than US personalities, even when it’s more technically advanced. That’s a branding challenge for the region, not a talent gap.
6. Looking ahead
What happens next depends less on gstack itself and more on how fast agentic development becomes normal.
Over the next 12–24 months, expect:
- Standardization of agent patterns. What Tan has published ad‑hoc will likely solidify into templates inside IDEs, cloud platforms and enterprise tools. "Spawn reviewer agent" or "run security agent pass" will become menu items rather than custom markdown files.
- Harder questions about accountability. If an AI-driven code reviewer "approves" a change that later causes a breach, whose fault is it? The engineer, the model provider, the person who designed the workflow? Regulators, especially in the EU, will not accept hand‑waving here.
- A split between hobbyist stacks and production‑grade systems. Gstack is optimized for one experienced user moving fast. Companies will need stricter controls: versioning of prompts, audit logs for agent decisions, and integration with existing CI/CD and security pipelines.
For individual developers and founders, the opportunity is to treat setups like gstack as starting points, not finished products. The real value comes from customizing them to your domain, your risk tolerance, and your regulatory reality.
The unresolved question is psychological: can we build agent workflows that enhance focus and creativity without recreating a new form of 24/7 hustle, where the "team" never stops working and neither does the human supervising it?
7. The bottom line
Garry Tan’s gstack is neither a revolution nor a joke; it’s a polished snapshot of where serious AI users already are. The hype around it exposes how much power celebrity still has in open source – and how quickly AI work is shifting from writing code to designing organizations of agents. For European builders, the smart move is to adopt the pattern, not the mythology: leverage AI teams, respect regulation, and refuse to romanticize sleeplessness as a business model.
