Y CombinatorHow to Build Superintelligence Inside Your Company
CHAPTERS
Building “superintelligence” as an organizational layer (not a copilot)
The hosts frame the episode around a shift from using AI as a helper to using it as foundational infrastructure. They preview the core idea: record organizational artifacts, make knowledge reusable, and let teams compound their effectiveness through shared agent systems.
YC’s internal agent stack: from a small harness to a company-wide platform
Pete describes how YC built an internal agent harness that started as a small engineering project and snowballed into shared infrastructure. The system includes agent loops, a model router, and a tool registry that teams across YC can extend.
The finance workflow pain that triggered everything
The original problem was the slow, inefficient loop between finance explaining complex workflows and engineers encoding them into deterministic software. Pete’s insight was to give finance the ability to control and encode workflows directly using prompts and agents.
The first real unlock: read-only SQL tools for non-technical teams
YC’s early “magic moment” came from letting agents query the production database (read-only), plus schema/model-file access. This turned natural language questions into real answers for non-technical users and made the system immediately valuable.
“One database to rule them all”: why unified context beats scattered SaaS
Because YC runs much of its operations on internally-built software, critical context lives in one Postgres schema. That centralization makes agents far more effective: they can answer arbitrary questions spanning companies, founders, notes, and transactions.
Jevons Paradox in analytics: lowering query cost increases curiosity
With agents writing SQL, the cost of asking questions collapses. That doesn’t just speed up answers—it changes behavior by dramatically increasing the number, scale, and ambition of questions people ask.
Denormalizing for agents: GBrain, retrieval, and “Bigtable for context”
The conversation shifts to companies that don’t have a single database of truth. The proposed workaround is deliberate “denormalization” into agent-optimized formats, supported by modern retrieval stacks (RAG/GraphRAG/hybrid reranking).
From single-player to multiplayer agents: the unsolved organizational harness
Pete argues most popular agent tools are designed for individual power on one machine. The harder, more valuable challenge is enabling “multiplayer” agents that work across teams with shared primitives, shared context, and shared governance.
350+ internal tools: the shared registry as the “work adapter layer”
YC’s tool registry evolved from ~20 tools to 350+ as teams added capabilities for their workflows. This registry is what makes agents actually useful at work, and it can also be reused by personal harnesses like Claude Code on employee machines.
Skillify, DRY/MECE resolvers, and the rise of “meta-skills”
Garry describes a pattern: turn repeated actions into skills, then keep the skill set clean via resolver checks that enforce DRY and MECE principles. The group observes these “registries/resolvers” emerging independently across different agent ecosystems.
The self-improving dream cycle: nightly review of conversations to get better
YC runs a general agent that reviews internal agent conversations to find improvements and missing context. This creates an automated feedback loop where the system learns from real usage and gradually becomes more capable and efficient.
The two-sentence pitch skill: capturing tacit partner expertise in prompts
They zoom into a concrete example: YC’s two-sentence company description. A partner-created skill improved significantly after feeding it meeting transcripts of founders practicing and receiving feedback, effectively distilling partner judgment into a reusable tool.
Recording everything: transcripts as a compounding “building layer”
Diana and Pete argue that recording meetings and producing artifacts used to feel socially awkward but is becoming normal because the ROI is massive. Those recordings can improve communications, planning, coaching, and system prompts—turning everyday work into reusable training material.
Shared organizational brain requires transparency + trust-default culture
YC made agent conversations broadly visible internally, enabling learning by observation and adding a social layer of accountability. The hosts argue that truly agentic organizations require egalitarian access and trust-by-default norms, plus a willingness to spend on tokens early to time-warp ahead.
Horseless Carriages, chat as the interface, and just-in-time software
Pete explains his critique of “AI-as-a-feature” software that hides prompts and keeps control with developers. The group argues the future is agent-first systems that wrap deterministic tools, with chat as the best interface and “just-in-time” software generated on demand instead of prebuilt UI-heavy apps.
Centralizing vs decentralizing AI: the coming personal AI revolution
The episode ends on a societal and organizational fork: AI could become centralized “mainframe-like” control or a personal-computing-style empowerment wave. They argue decentralization requires deliberate choices—open prompts, user-owned systems, and broad access—rather than default corporate lock-down.
