How To Build Superintelligence Inside Your Company
YC General Partner Pete Koomen describes how Y Combinator built an internal AI agent infrastructure over the past year, transforming their organization through shared tool registries, SQL access to a unified database, and self-improving skill systems. The conversation explores how this approach represents a blueprint for building 'superintelligence' inside any organization, contrasting open, trust-based AI adoption against centralized corporate control of AI tools.
Summary
The episode features YC General Partner Pete Koomen discussing how Y Combinator built its internal AI agent infrastructure, starting approximately a year ago from a specific project with the finance team. The original impetus was to eliminate the inefficient loop between finance staff and software engineers by giving non-technical employees the ability to encode their own workflows in natural language rather than code. The first major unlock was giving agents read-only SQL access to YC's single Postgres database, which contains every important piece of organizational context — from funded companies and founders to financial transactions and CRM notes. This dramatically increased both the volume and complexity of questions employees were willing to ask, illustrating Jevons paradox applied to organizational intelligence.
The infrastructure evolved from a simple agent loop and tool registry into a system with over 350 tools, contributed by teams across the organization. A key concept discussed is the 'skill registry' — an abstraction layer over tools that encodes organizational knowledge as prompts. A concrete example is the 'two-sentence description' skill used to help founders articulate what their company does. This skill was initially hand-written by a partner, then improved by feeding in transcripts of group office hours sessions where founders practiced the skill and received feedback. An automated nightly agent reads all conversation transcripts and identifies opportunities to improve skills, creating a self-improving loop the hosts compare to Karpathy's 'dream cycle.'
The conversation addresses the 'multiplayer' challenge in AI — most current agent harnesses (Claude Code, OpenClaw, Hermes) are single-user tools, but the real organizational opportunity lies in making these capabilities available at a team and company level. YC's approach defaulted all agent conversations to be publicly viewable by all full-time employees via a Slack broadcast channel. This transparency served multiple purposes: employees learned how to use the tools by watching colleagues, it instituted social controls on sensitive information, and it allowed new employees to quickly apprentice off the best performers in the organization by accessing their recorded interactions.
The hosts frame this as a choice between two futures for AI: a centralized model where a handful of large corporations control AI capabilities and users cannot modify their own prompts (analogized to the mainframe era), versus a decentralized personal AI revolution analogized to the Apple I and the homebrew computer club. Tools like OpenClaw, Hermes Agent, and Gbrain are positioned as representing the latter path — where individuals control their own prompts, choose their own models, connect their own databases, and extend their own software. The episode also critiques 'horseless carriage' AI software design, where AI is inserted as a small feature inside largely deterministic software, arguing the future belongs to minimal harnesses where agents wrap tools rather than software wrapping AI. Chat is defended as the right interface because it most closely maps to human language and thought, enabling just-in-time software generation rather than rigid pre-built UIs.
Key Insights
- Pete Koomen argues that giving agents unrestricted read-only SQL access to a single unified database was the core unlock — because YC runs all operations on one Postgres schema rather than fragmented SaaS tools, agents could answer arbitrarily complex business questions without human data science intermediaries.
- The hosts observe that removing friction from asking questions dramatically increased the number and complexity of questions asked — an application of Jevons paradox where cheaper access to answers creates more demand for answers, not less.
- Koomen describes a self-improving skill loop where a nightly agent reads all employee-agent conversation transcripts, identifies gaps in existing skills, and automatically updates them — analogized to Karpathy's 'dream cycle' concept.
- Gary argues that building organizational superintelligence is not technically complex — it is the aggregate result of encoding every atomic organizational skill as a prompt, feeding artifacts and transcripts back into those skills, and running automated improvement loops on all of them simultaneously.
- The hosts contend that truly agentic organizations require two cultural preconditions most companies lack: egalitarian access (not limiting AI to leadership) and trust-by-default (defaulting to open rather than locked-down data access), which is why this model works best at startups.
- Koomen's 'horseless carriage' critique holds that most AI software keeps prompt context locked away from users inside deterministic software wrappers, which is the inverse of the correct architecture — the future is minimal harnesses where agents wrap deterministic tools, not software that wraps AI as a feature.
- Gary frames the current moment as a fork between two futures: a centralized AI world where five corporations control all compute and users cannot modify their own prompts (analogized to the mainframe era), versus a personal AI revolution where individuals control models, prompts, and data (analogized to the Apple I and homebrew computing).
- The hosts observe that defaulting all internal agent conversations to public broadcast within YC simultaneously solved three problems: it accelerated onboarding by letting new employees learn from watching expert usage, it created social controls that kept sensitive information private without technical enforcement, and it allowed the organization to identify and replicate the behaviors of its best performers at scale.
Topics
Full transcript available for MurmurCast members
Sign Up to Access