How to Build Superintelligence Inside Your Company
YC General Partner Pete Kumman describes how Y Combinator built internal AI agent infrastructure over the past year, including a shared tool registry with 350+ tools, a centralized database for agent context, and self-improving skill loops. The conversation covers the principles behind building 'superintelligence' inside organizations through shared context, transparency, and agentic workflows that compound organizational knowledge over time.
Summary
The episode features YC General Partner Pete Kumman discussing how Y Combinator transformed itself into an AI-native organization by building internal agent infrastructure from scratch. The journey began about a year prior when Pete and a small engineering team worked with YC's finance team, noticing an inefficient loop where software engineers had to manually translate complex financial workflows into deterministic software. The realization that agentic coding tools could empower non-technical employees to encode their own workflows in plain English prompted a broader infrastructure project.
The foundational breakthrough came when agents were given read-only SQL access to YC's single PostgreSQL database, which houses all critical company context — funded companies, founders, financial transactions, CRM notes, and more. This single-database architecture proved enormously powerful because it allowed agents to answer arbitrary, complex questions instantly, dramatically increasing both the volume and complexity of questions employees would dare to ask — an instance of Jevons' paradox applied to organizational intelligence.
The team built a shared internal tool registry that has grown from roughly 20 tools to over 350, covering finance workflows, office hours management, event planning, and more. Skills — abstraction layers over tools — emerged as a key primitive, and the system evolved toward self-improvement: a nightly agent reads all employee agent conversations, identifies gaps, and automatically improves skills. A concrete example is YC's 'two-sentence description' skill, which was refined by feeding it transcripts of partner office hours sessions, ultimately making the AI better at this task than any individual partner.
A critical organizational design choice was making all agent conversations publicly viewable by full-time YC employees via a Slack broadcast channel. This transparency served multiple purposes: it allowed employees to learn from each other's agent usage, it acted as a social control mechanism to keep sensitive information private without heavy technical restrictions, and it enabled a high-trust, egalitarian environment that the hosts argue is prerequisite for this type of AI-native organization.
The hosts draw broader lessons about the nature of AI adoption in organizations. They argue that the 'horseless carriage' problem — slotting AI as a small feature inside deterministic software rather than building agent-first — still dominates commercial software. They advocate for minimal, agent-wrapping architectures ('just-in-time software') over large deterministic codebases, citing Gary's experience rewriting a half-million-line Rails app into a far smaller, more flexible agentic system. The conversation concludes with a philosophical argument about centralization versus decentralization in AI: the hosts warn against a 'mainframe era' scenario where a few large companies control AI and user prompts, and advocate for a personal computing-style revolution where individuals own and program their own AI systems.
Key Insights
- Pete Kumman argues that giving agents unrestricted read-only access to YC's single production PostgreSQL database — despite initial security concerns — was the key unlock that allowed non-technical employees to ask arbitrarily complex business questions instantly, dramatically increasing both the frequency and ambition of questions people would ask.
- Kumman describes how YC's 'two-sentence description' skill was automatically improved by feeding an agent transcripts of partner office hours sessions, with partners reporting the skill became noticeably better than any individual partner at writing these descriptions — a concrete micro-mechanism for how organizational superintelligence compounds.
- The hosts argue that defaulting all internal agent conversations to a public Slack broadcast channel simultaneously solved the learning problem (employees watched each other's usage and learned new techniques), the security problem (social transparency acted as a control), and the ramp-up problem (new employees could apprentice against the AI embodiment of top performers).
- Gary contends that building a superintelligent organization requires two cultural prerequisites that most organizations lack by default: being egalitarian (giving all staff, not just leadership, access to these tools) and being trust-by-default (rather than locking down context for security reasons), arguing these traits are more natural to startups than incumbents.
- Pete Kumman's 'Horseless Carriages' essay argues that most commercial AI software keeps prompt context locked away from users — a developer-controlled model — whereas the true potential of AI is to shift control of software from developers to users, and that future AI-native software will look like agents wrapping deterministic tools rather than deterministic software wrapping AI.
Topics
Full transcript available for MurmurCast members
Sign Up to Access