Y Combinator argues founder-owned AI agents will shrink startup teams

Why it matters

YC is telling founders to treat their private context and operating procedures as core infrastructure. That could reduce early hiring while creating new security and ownership disputes around executable workplace knowledge.

Y Combinator on August 6th published a Startup School 2026 talk in which Garry Tan (@garrytan) argued that founders can build companies with fewer employees by giving AI agents persistent memory, reusable operating instructions and access to their own infrastructure.

Tan calls the setup "personal AGI." The phrase does not describe a new foundation model or a measured advance toward artificial general intelligence. It is Tan's name for a founder-controlled operating layer that connects existing models to private context and repeatable workflows.

His central claim is that access to frontier models will become broadly available, leaving a founder's accumulated knowledge as the scarce asset. A general-purpose chatbot starts with roughly the same capabilities for every subscriber. An agent connected to years of emails, meetings, decisions and working procedures can act with context that competitors cannot buy from the model provider.

"Personal AGI is how you stay under your own power in the age of agents," Tan said in the full transcript of the talk.

Tan is Y Combinator's president and CEO, a role that gives his operating prescription added weight among early-stage founders. He previously co-founded the blogging platform Posterous, worked as an early designer and engineering manager at Palantir, served as a YC partner and co-founded Initialized Capital before returning to run the accelerator.

The personal AGI stack

Tan described his system as three connected components: a frontier model rented from an AI provider, a library of personal context controlled by the user and a software harness that lets agents retrieve information and perform work.

The memory layer is GBrain, Tan's open-source knowledge system for AI agents. Tan said his installation contains about 220,000 markdown pages covering 25 years of emails, meetings, notes, photographs, drafts and decisions. Agents compile, organize and search that material, according to Tan, producing meeting briefs, inbox triage and overnight research.

Above that memory sits gstack, a collection of agent workflows for software development and company operations. Tan said during the talk that the repository had 123,000 GitHub stars. GitHub showed roughly 125,000 when checked around the publication of the talk.

The workflows are largely written as "skill files": plain-English instructions stored in markdown. One example tells an agent how to process a meeting transcript, identify commitments and deadlines, cross-reference the people involved and flag contradictions without overwriting earlier records.

Tan's argument stretches beyond coding. He said YC employees in media, events and finance are writing skill files and scheduled agent jobs, including a finance employee who used an internal agent to turn about 100 Excel workbooks into an application. Those examples are YC's own account of its internal use and do not include independent measurements of time saved or output quality.

A pitch for smaller companies

Tan tied the stack directly to startup headcount. A founder can begin with an organization made up of one person and a set of agents, he said, assigning recurring jobs before hiring employees or incorporating a company.

He calculated that his own coding output had risen as much as 400 times from his 2013 baseline, while acknowledging that lines of code are an imperfect productivity measure. After applying his own penalty for scaffolding, bloated output and overstatement, Tan placed the lower bound at eight times. The calculation remains self-reported.

Tan also said a quarter of YC's Winter 2025 companies had codebases that were 95% AI-generated and described that group as one of YC's fastest-growing and most profitable batches. He stopped short of claiming that AI-generated code caused those results.

That distinction matters. A smaller engineering team can ship more code with agents while accumulating review, maintenance and security work that raw output does not capture. Tan's framework addresses part of that problem by separating model judgment from deterministic operations such as arithmetic, database queries and scheduling. His agent writes or calls conventional code when a task requires exact results.

The prescription for founders is practical: run an agent on infrastructure they control, assemble an initial library from notes and email, document one recurring task as a skill, schedule it and preserve the resulting workflow for reuse. Tan said users should expect a rough first month before the stored context and procedures begin producing useful compounding returns.

YC benefits if that thesis proves correct. Lower staffing requirements reduce the capital and recruiting needed to test a startup, expanding the number of founders who can reach a working product. The same shift creates fundable markets for agent infrastructure, memory systems, security, deployment and collaboration tools.

Ownership creates a new fault line

Tan's sharper argument concerns who owns the instructions that encode a worker's judgment. A support engineer might teach an agent how to triage incidents, calm a customer or write a postmortem. Those procedures could remain in the worker's repository and move with them, or stay inside an employer's systems after the worker leaves.

"Own your skills because if you don't, your job becomes a skill file," Tan said.

That position sets up a conflict between personal portability and a company's interest in retaining processes developed during employment. As agent workflows become executable operating knowledge, founders will need explicit rules for which skills belong to an individual, which belong to the company and what can leave with an employee.

The same ownership thesis carries a security burden. Tan's system consolidates email, meetings, personal history and company information into a repository that becomes an unusually valuable target. He argues that running it under the user's keys and infrastructure provides better custody than scattering the information across cloud services. That custody also leaves the user responsible for access controls, backups, software updates and incident response.

Tan has open-sourced his own tools and said founders can use other agent harnesses or coding products. His strategic point is independent of any single repository: model access is becoming interchangeable, while proprietary context and well-maintained procedures can accumulate into a durable operating advantage.

If that advantage holds, the smallest viable startup team will keep shrinking. The founder's job will shift toward defining processes, reviewing agent output and deciding which knowledge deserves to become reusable software.