Garry Tan on AI Startups: Stop Chasing Heat and Build Reusable Loops

Original video: Garry Tan: New Rules for Founders

a16z · 2026-08-12 · 51 minutes 28 seconds

Interviewer: Anish Acharya, General Partner at a16z · Guest: Garry Tan, President and CEO of Y Combinator

Additional primary sources: Y Combinator profile · Garry Tan's gstack repository

In 2003, Silicon Valley offered few jobs after the dot-com crash. Garry Tan was 22 and had two choices, one at Expedia and one working on Windows Mobile at Microsoft. He chose the mobile field that appeared to have more momentum and left web development, where he had already accumulated years of experience.

Around the same time, Peter Thiel and Tan's Stanford friends invited him to join the team that would become Palantir. Thiel even offered a 70,000 dollar check, equal to Tan's Microsoft salary. Tan declined because a promotion at Microsoft looked possible.

More than two decades later, he places both decisions in the same category. He worked backward from an external heat map and ignored the terrain that he could already see directly.

That experience provides the foundation for the interview. AI is reducing the cost of implementing software, allowing founders to do more with fewer people. As the supply of code increases, the larger advantages move toward choosing the right problem, holding direct information, and turning one success into a loop that keeps running.

Stop working backward from the heat map

The prevailing message in 2003 was that the web had ended and mobile would come next. The timing was almost exactly wrong. Social networks were about to begin another platform cycle. Tan's existing web experience could have been an advantage, yet he treated it as an obsolete asset.

His Palantir decision had the same structure. In retrospect, the strongest information was concrete. The invitation came from some of the most capable people he knew, and they were bringing Silicon Valley software into government and large institutions. Job titles, levels, and fashionable categories on the career map overwhelmed those close-range signals.

Tan therefore places direct experience at the beginning of company formation. A problem may keep hurting a specific group, a technology may already have a real use inside a small community, or an exceptional team may be building something that outsiders do not yet understand. Those signals are often closer to an opportunity than a list of what is currently hot.

YC serves a similar role in his account. An online application lowered the social barrier to Silicon Valley. The founder community then gave members a place to discuss losing a customer, an engineer quitting, or a co-founder losing hope. Public success narratives are easy to distort. Close-range information that can be questioned is more useful.

As code gets cheaper, taste and agency get more valuable

Tan says vibe coding and coding agents can turn one person into 400x their former self. The phrase communicates scale. The interview supplies no reproducible productivity test, so it should not be read as a benchmark.

The directional change is still clear. A founder once had to write a specification, schedule engineering time, arrange testing, and wait for delivery. Many low-risk projects can now be assigned directly to an agent, leaving more founder time for framing the problem, judging the result, and choosing the next iteration.

Easier software generation also puts pressure on pure per-seat SaaS. Tan's view in 2026 is that a product in this category needs a route toward data, network effects, or another hard-to-copy asset if it expects to preserve an advantage over the next 5 to 10 years. This is his product and investment thesis. The market has not produced one settled answer.

Tan rebuilt his own coding practice through gstack. He tried models on low-stakes ideas, then organized product, engineering, testing, and browser acceptance into reusable Skills. One prompt derived from YC office-hour recordings was too intense, so he had a model reduce its strength by 90% before releasing it.

That detail matters. An agent can produce quickly, while the founder still decides what to keep, remove, or soften. Taste and agency do not disappear with automation. Faster implementation puts them under review more often.

Doing a task once is not enough, turn it into a loop

The most concrete method in the interview is to convert business processes into Skills.

A person and an agent first complete a task together. The useful steps then become Markdown instructions, accompanied by any required code and tests, plus a trigger and acceptance criteria. When the same kind of task returns, the system reuses that process. If it fails, the correction is written back into the Skill and every later run carries the lesson.

Tan condenses the idea into the phrase “a Markdown file is an employee.” A more precise description is a repeatable work definition that can keep accumulating fixes. Employees still carry judgment, collaboration, and responsibility. The file preserves the process and the part a machine can execute.

The QA loop in gstack offers an engineering example. Once product and code generation are automated, the bottleneck moves to black-box acceptance. The system must open a browser, complete the flow, find the mismatch, and carry the fix into the next run. The closed feedback loop creates the leverage, along with the speed of the first generation.

Tan also says he has seen teams made up of 2 to 3 people and hundreds of Skill files go from zero to 15M dollars in ARR within 4 months. The interview provides no company name, measurement definition, or independent material. The figure is an attributed anecdote, not a general growth benchmark for small teams.

Token maxing is a preview of a future workflow

Tan calls his high-intensity use of agents token maxing. He lets a system read 800,000 to 1,000,000 tokens for a single task and estimates that running at full strength can cost 50,000 to 100,000 dollars per year. He describes the experience as living in 2028 today.

The estimate does not specify a model mix, request volume, or billing record. It should be treated as a personal approximation. The unit economics are also beyond what most products can currently support.

The reusable lesson comes from using a high-capability system to expose business bottlenecks early. A founder observes where the strongest available agent fails on real work, then turns the solution into a Skill, code, and tests. Compute pays for exploration. The process is the asset that remains.

More files create a new debt, provenance and conflict

As the number of Skills grows, a system encounters another class of problem. The same fact may have two versions, an old rule may conflict with a new one, and a conclusion may come from an unverified meeting remark.

Tan says the system needs provenance and time. It should know which information is newer, who supplied it, why it should replace an older version, and how periodic correction will work. Without those records, long-term memory becomes a larger collection of text that is harder to judge.

He also describes an example from Brex management. An agent read the meeting records of direct reports, giving a leader roughly 3 weeks of context across 2 levels of the organization and allowing quick intervention in a conflict. The source package contains no independent material verifying the anecdote.

From a product-governance perspective, the example also exposes an authority boundary. Meeting records contain employee, customer, and compliance information. Without informed authorization, minimal access, retention limits, and audit records, stronger retrieval can slide into organizational surveillance. Retrieval quality is only one part of a safe operating design, which also needs explicit authority and accountability. The ability to read everything does not confer the right to use everything.

A startup can redraw the work instead of only speeding up the old process

Tan recalls that his Windows Mobile team at Microsoft was blocked by a dependency on another division. The other group neither fixed the bug nor clearly refused it. Two technical employees eventually had to cross the campus and chase the issue in person. Engineering capability was not the problem. Information and responsibility were trapped between organizational layers.

He uses the familiar 7±2 shorthand for the limits of human attention. The figure is a simplified version of working-memory research, but it points to a real organizational constraint. When a business no longer fits inside one person's head, email, meetings, and middle-management reporting begin to synchronize it, losing information at every handoff.

Agents can continuously read dependencies, test results, customer feedback, and meeting records. That creates room to redesign the organizational layer. Product leaders set direction, people close to the work retain execution authority, and agents organize status, surface blockers, and run repeated checks.

Large companies cannot replace their existing structure in one move. A startup can design a different operating model from its first day. That advantage has conditions. People must still define objectives, handle exceptions, approve high-risk actions, and own the outcome. Automation expands execution bandwidth and can expand the reach of a bad permission or bad objective at the same time.

Models move fast, organizations change slowly

Tan's forecast for AI adoption is comparatively slow. He sees people aged 18 to 22 at Startup School treating AI as a natural tool, much as the previous founder generation treated the web and mobile devices as infrastructure.

It takes time for that cohort to enter and run companies, institutions, and governments. His rough horizon is 20 years. Models may change every few months. Budgets, compliance rules, roles, procurement systems, and trust relationships will not be rewritten on the same schedule.

This is the source of his optimistic case. Large-scale job displacement will not follow the cadence of model releases instantaneously, leaving people and organizations an adaptation window. Companies such as Microsoft also possess customer relationships, distribution, data, and institutional positions that do not vanish when a stronger model appears.

The opportunity for startups comes from the speed gap. They can establish a closed loop inside one concrete workflow, prove quality, authority, and economics, then expand. Attaching a new model to the end of an old process rarely creates the same structural advantage.

The next computer may remember people before it recedes behind the interface

On consumer AI, Tan expects the hardware form factor to remain recognizable in the short term while interaction shifts toward voice, memory, computer use, and persistent context. The system would understand a person's goals, preferences, and history, then search continuously for useful ways to help.

He predicts a round of “harness wars” in 2027. Competition would move beyond one model response toward the complete operating environment, including context, memory, tools, permissions, and continuity across devices. Both the label and the year are his forecast.

Cost remains a constraint. Tan estimates that a workload requiring frontier compute today could fall into a 50 to 100 dollar consumer range within 2 to 3 years. The interview does not define that workload, and pricing will move with models and products. The direction is more defensible than the exact figure. Lower costs give consumer software room for free trials, long-lived memory, and frequent actions.

Technology can compress execution, not responsibility

The final part of the interview turns to local government in San Francisco. Tan connects it to the agent discussion through one conclusion: technology can accelerate information processing and execution, while social coordination still requires people to speak publicly, organize action, and accept reputational cost.

The same boundary applies inside a company. An agent can write code, organize meetings, find conflicts, and execute processes. It cannot decide on its own what deserves to exist, acquire a legitimate right to read every record, or absorb the consequences of a management decision.

The interview's new rules for founders can be reduced to four:

1. Start from direct experience and unusual knowledge, with less dependence on a heat map of popular projects.

2. Turn successful tasks into loops with code, tests, feedback, and a correction history.

3. Establish provenance, authority, conflict, and audit rules before Skills and memories multiply.

4. Separate model speed from organizational speed, keep forecasts labeled as forecasts, and leave responsibility with people.

AI expands implementation capacity quickly and exposes choices and governance earlier. Small teams have a large opening if their loops can run fast while still knowing when to stop, who must confirm the result, and how a reliable correction survives the next failure.