The atmosphere at South by Southwest (SXSW) is traditionally one of cautious optimism and creative exploration, but during a recent onstage conversation with venture capital titan Bill Gurley, Y Combinator CEO Garry Tan introduced a more frenetic energy. Tan, a figure synonymous with the high-stakes world of Silicon Valley incubation, confessed to the audience that he is currently operating in a state he describes as "cyber psychosis." Driven by a relentless obsession with the latest evolution in artificial intelligence—AI agents—Tan revealed he has been surviving on a mere four hours of sleep per night. This admission was not merely a flourish of "hustle culture" bravado; it served as the preamble to the release of "gstack," an open-source configuration for Anthropic’s Claude Code that has since ignited a firestorm of debate across the global developer community.
To understand the fervor surrounding gstack, one must first understand the shift from generative AI as a passive assistant to AI as an active agent. While tools like GitHub Copilot have long offered "autocomplete" for code, the new frontier involves agentic workflows where the AI can execute terminal commands, run tests, and iterate on complex multi-file architectures. Tan’s contribution to this space is not a new Large Language Model (LLM), but rather a sophisticated orchestration layer—a collection of "opinionated" skills that dictate how an AI should behave when navigating a codebase.
The genesis of gstack lies in Tan’s own history as a founder. During the interview, he drew a vivid parallel between his current AI-driven workflow and the early, grueling days of Posterous, the blogging startup he co-founded and eventually sold to Twitter in 2012. He recalled the era of $10 million in venture capital and a ten-person team, a period so demanding he relied on modafinil, a potent wakefulness-promoting agent, to maintain the pace. Today, Tan claims the pharmacological assistance is unnecessary. The "natural insomnia" provided by the "AI revolution" has replaced the need for stimulants, as he manages what he describes as ten virtual workers across three simultaneous projects.
Released on GitHub under an MIT license, gstack is essentially a repository of "skills"—reusable prompt instructions stored in .md files that Claude Code utilizes to adopt specific personas and methodologies. These skills range from a "CEO" persona that evaluates the strategic viability of a feature, to an "Engineer" that writes the implementation, and a "Security Reviewer" that audits the output for vulnerabilities. By chaining these roles together, Tan has created a digital microcosm of a traditional engineering organization, allowing a single human operator to act as the conductor of a high-speed, multi-disciplinary symphony.
The reception to gstack was immediate and divided. On one hand, the project exploded in popularity, trending on Product Hunt and amassing nearly 20,000 stars on GitHub within days. For many, gstack represents a "pro-level" configuration that democratizes the complex prompt engineering required to make AI agents truly productive. Proponents argue that the value of gstack is not in the individual lines of text, but in the structural insight it provides: that AI performs best when it is forced to simulate the checks and balances of a human hierarchy.
However, the "love" for the project was quickly met with a "heap of hate." The controversy reached a boiling point when Tan shared a testimonial from a CTO friend who claimed gstack was in "god mode," allegedly discovering a subtle cross-site scripting (XSS) vulnerability that a human team had missed. The friend went so far as to predict that 90% of all new code repositories would eventually use gstack. This sparked a backlash from the developer community, with critics labeling the rhetoric as delusional. Some argued that any competent CTO whose team missed a basic XSS flaw should be terminated, while others pointed out that gstack is, at its core, a collection of text prompts that many senior developers had already internalized or automated themselves.
The criticism often touched on the "celebrity" nature of the release. Commenters on Product Hunt and X (formerly Twitter) suggested that the project’s viral success was a byproduct of Tan’s status as the head of Y Combinator rather than any inherent technical breakthrough. To these skeptics, gstack is a "wrapper" of existing functionality, a symptom of an industry where the hype surrounding AI agents has outpaced the actual utility of the tools.
To bridge the gap between the hyperbole and the skepticism, an analysis of the technical foundations of gstack is required. When queried for an objective assessment, the industry’s leading LLMs—ChatGPT, Gemini, and Claude—offered surprisingly nuanced perspectives. ChatGPT noted that while the workflows are not "magical," the real insight is the simulation of an engineering organizational structure. Gemini characterized it as a "Pro" configuration focused on "correctness" over mere ease of use. Claude, perhaps predictably, praised the system as a mature, opinionated framework built by a power user.
From a technical standpoint, the efficacy of gstack lies in its "role-playing" architecture. Research into LLM performance has consistently shown that "Chain of Thought" prompting and persona-based instruction significantly reduce hallucination rates and improve logic. By forcing the AI to first "think" like a CEO and then "review" like a security lead, gstack implements a heuristic for quality control that a single, broad prompt cannot achieve. This "multi-agent" approach is increasingly seen as the standard for high-fidelity AI development.
Beyond the code, the gstack saga highlights a shifting paradigm in the tech industry: the "One-Person Unicorn." Tan’s excitement stems from the belief that the barrier to entry for building complex software is collapsing. If a single founder can use gstack to replicate the output of a ten-person team that previously required $10 million in funding, the venture capital model itself may face a fundamental reckoning. If capital is no longer the primary bottleneck for scaling labor, the value of a firm like Y Combinator shifts from providing "money for hiring" to providing "access to the best agentic stacks."
However, this transition is not without its psychological and societal costs. Tan’s description of "cyber psychosis" and his return to a state of near-sleeplessness reflects a broader trend of AI-induced burnout. The "momentary crystalline structures" of code that Tan describes chasing into the early morning hours suggest a new form of digital obsession, where the speed of AI creation outruns the human capacity for rest and reflection. As AI agents become more capable, the pressure on human founders to remain "always on" to manage their digital workforce may create a mental health crisis that the industry is currently ill-equipped to handle.
Looking toward the future, the legacy of gstack will likely be its role as a precursor to more integrated "Personal Engineering Stacks." We are moving toward a world where every developer and founder will maintain a customized "skill library"—a digital DNA that defines their unique approach to problem-solving. Whether gstack itself remains the dominant configuration is secondary to the fact that it has normalized the idea of "opinionated" AI.
As the industry prepares for major milestones, such as the upcoming technological summits in San Francisco in late 2026, the debate over gstack serves as a bellwether for the "Agentic Age." The polarization it caused is a reflection of the industry’s anxiety over its own automation. For some, Tan’s setup is a tool for liberation, a way to reclaim the raw creative power of the early internet. For others, it is a source of "delusional" hype that obscures the hard, human work of engineering.
In the final analysis, Garry Tan’s gstack is more than just a GitHub repository; it is a manifesto for the next decade of software development. It posits that the most valuable asset in the AI era is not the model itself, but the "opinion" of the person directing it. As Tan himself articulated, the experience of speaking to a machine and watching a structure build itself is a power that, once tasted, is difficult to abandon. Whether that power leads to a new era of innovation or a widespread case of "cyber psychosis" remains to be seen, but the era of the passive AI assistant is officially over. The agents have arrived, and they are following Garry Tan’s instructions.