Interview
Inside YC's AI Playbook
- YC has transitioned from a traditional pre-AI organization to a "super AI-native" entity by treating AI as a foundational building layer rather than merely a co-pilot.
- The initiative began approximately one year ago as a specific project for the finance team to encode complex workflows using English prompts instead of deterministic code (e.g., Ruby).
- YC runs on proprietary software hosted on a single PostgreSQL database containing all critical context (funded companies, financial transactions, internal notes), which serves as a unified "shared organizational brain."
- A key architectural breakthrough was granting agents unrestricted read-only access to this production database, enabling them to answer arbitrary business questions that previously required hours of manual SQL coding.
- YC has built an internal infrastructure with over 350 specific tools, including a tool registry (or "resolver") that maps agent skills to underlying functions, allowing for dynamic, self-improving workflows.
- The organization utilizes a "dream cycle" process where an autonomous agent analyzes meeting transcripts and agent interactions daily to identify inefficiencies and automatically refine or create new skills.
- A specific example of organizational superintelligence is the automated "two-sentence description" skill, which has improved beyond human capability by ingesting transcripts of successful founder feedback sessions to refine its prompting logic.
- YC operates on a default transparency model where all internal agent conversations are broadcast to public Slack channels, leveraging social control and peer learning to manage security risks.
- Pete Kuman argues that building such "superintelligent" organizations requires a high-trust, egalitarian culture and a willingness to spend $10,000–$100,000 annually on AI tokens to leverage context that will eventually become commoditized.
- The "Horseless Carriages" essay critiques the industry trend of embedding AI as a hidden feature within traditional software, advocating instead for "just-in-time" software where the chat interface is the primary UI and the user controls the prompts.
- The discussion identifies a divergence in the industry's future between a centralized "1984" model of mainframe-like control and a decentralized "Apple I" model where individuals own their agents, prompts, and data repositories.
- YC's internal success is framed as a "one-time time warp" that allows them to leapfrog incumbents by operating with a 2028 mindset of open, autonomous agents while competitors remain stuck in 2023-style co-pilot usage.
- The shift enables a "raising the floor" effect where new employees achieve rapid ramp-up times by accessing the aggregated context and skills of star performers via AI agents, rather than relying on scarce human mentorship.