Garry Tan's gstack wants to make AI coding agents think before they ship

Garry Tan, the president and CEO of Y Combinator, has released gstack, an open-source toolkit that wraps Anthropic's Claude Code inside eight distinct workflow modes covering product planning, engineering review, quality assurance, and release management. The project is available publicly and positions itself as an opinionated system for teams and solo developers who want AI-assisted coding to behave less like an eager intern and more like a disciplined engineering organization.

The core premise is deceptively simple: most AI coding tools collapse the entire software development lifecycle into a single conversational thread, which means the model is simultaneously trying to understand requirements, write code, review its own work, and decide when something is ready to ship. That is roughly equivalent to asking one person to be the product manager, lead engineer, QA tester, and release manager all at once, without ever switching hats. Errors compound. Context bleeds. The result is code that technically runs but carries the structural debt of decisions made too quickly.

gstack addresses this by enforcing separation of concerns at the workflow level. Each of its eight skills is designed to operate in a distinct mode, so the planning phase is not contaminated by implementation pressure, and the review phase is not skipped because the generation phase felt confident. The toolkit also runs on a persistent browser runtime, which means state is maintained across sessions rather than evaporating at the end of a conversation window.

Tan's decision to release this as open-source is worth examining on its own terms. Y Combinator has a direct institutional interest in the productivity of early-stage founders, many of whom are now building software with AI assistance from day one. If the tools those founders rely on produce brittle, poorly reviewed code, the downstream costs show up as technical debt, security vulnerabilities, and failed products. gstack is, in that light, as much an infrastructure investment in the YC ecosystem as it is a personal project.

There is also a broader signal here about where the AI coding market is heading. The first wave of AI coding tools competed on raw generation speed and the apparent magic of watching code appear from a prompt. The second wave, which gstack represents, is competing on reliability, process, and the kind of trust that comes from knowing a system will not skip the unglamorous parts of software development. Anthropic's Claude Code is already positioned as a more careful, instruction-following model compared to some competitors, and gstack layers an additional discipline framework on top of that foundation.

The persistent browser runtime is a detail that deserves more attention than it typically receives in coverage of tools like this. Stateless AI interactions are one of the quiet failure modes of current developer tooling. When context resets between sessions, the model loses the accumulated understanding of a codebase's quirks, the decisions that were made and why, and the half-finished reasoning from a previous work session. Persistence does not solve hallucination, but it does reduce the category of errors that come from the model not knowing what it already agreed to.

If gstack or tools like it gain meaningful adoption, the second-order effect most worth watching is what happens to the implicit norms of software quality in AI-assisted teams. Right now, there is a quiet race to the bottom in some corners of the startup world, where the speed of AI generation has outpaced the cultural infrastructure for reviewing what gets generated. Teams ship faster but review less, and the gap between what the code does and what the team thinks it does quietly widens.

A toolkit that enforces review and QA as non-optional workflow stages could help reverse that dynamic, not by slowing teams down, but by making the responsible path the path of least resistance. That is how good tooling actually changes behavior: not through warnings or documentation, but by making the right thing the easy thing.

The deeper question is whether the eight-skill structure Tan has chosen reflects a genuinely universal model of software development or a particular philosophy shaped by YC's portfolio companies and their specific needs. Open-source release invites that conversation, and the forks and contributions that follow will be as revealing as the original design.

What gstack ultimately represents is a bet that the next competitive advantage in AI-assisted development will not come from a faster model or a more impressive demo, but from the unglamorous work of building systems that know when to slow down.