From AI Features to AI Operating System: Building Superintelligence

The Y Combinator Podcast reveals a profound shift in organizational intelligence: moving beyond AI as a mere feature to embracing it as the fundamental operating system. This conversation with Pete Koomen unpacks how YC transformed its internal operations by treating AI not as a co-pilot, but as a "shared organizational brain," enabling unprecedented collective skill enhancement. The implications are vast, suggesting a future where organizations achieve superintelligence not through incremental AI adoption, but by fundamentally re-architecting their workflows and data structures around AI agents. This insight is crucial for leaders aiming to build truly AI-native organizations and gain a significant competitive edge by leveraging collective intelligence in ways previously unimaginable.

The Unseen Architecture: From AI Features to an AI Operating System

The conventional wisdom around AI adoption often focuses on integrating AI features into existing software -- a chatbot here, an automated report there. However, Pete Koomen's experience leading YC's internal AI infrastructure paints a starkly different picture. The real transformation, he argues, comes from treating AI as the fundamental building layer, the operating system upon which the entire organization runs. This isn't about adding AI as a feature; it's about making AI the foundation. The immediate benefit might seem like enhanced efficiency, but the downstream consequences are far more profound: the creation of a "shared organizational brain" that amplifies collective intelligence and skill.

This shift requires a radical departure from traditional software development and data management. Koomen highlights how YC's internal finance team, initially reliant on a slow, iterative process of describing workflows to engineers who then built purpose-built software, was revolutionized by giving agents direct access to their data. The key unlock? A single, unified Postgres database containing all of YC's critical information -- companies funded, founder details, financial transactions, internal notes. This consolidation, a deliberate act of denormalization for agentic access, allowed AI agents to answer complex, arbitrary questions about the business with unprecedented speed and scale.

"It just turns out when all of that context is in one place, with a little bit of additional information about how the schema is laid out, an agent can go and answer arbitrary questions about our business. That was a magic moment for sure when I first saw that."

This unification directly combats the inefficiencies of fragmented data, a common ailment in legacy organizations. Koomen likens this to Jevons paradox: when the effort required to access information decreases dramatically, the number and complexity of questions asked skyrocket. What once took hours of manual SQL writing and inter-team coordination became an instant query. This isn't just about faster answers; it's about unlocking a new level of inquiry and understanding that was previously impractical, creating a significant competitive advantage for those willing to consolidate their data.

The Unforeseen Power of Unrestricted Access

A critical, and perhaps counter-intuitive, insight from YC's journey is the power of giving agents broad, even unrestricted, access to data. Koomen recounts how Jared, a software engineer, initially felt he was "breaking the rules" by granting agents read-only SQL access to the production database. The fear of security breaches and data misuse is a pervasive concern, leading most organizations to impose strict limitations. However, Koomen observes that this "worrying a bit less" about security, within a high-trust environment, revealed the agents' "unbelievable power."

This contrasts sharply with the "single-player era" of agents, where tools like ChatGPT or Claude Code are designed for individual use, creating personal superpowers. The true organizational leap, Koomen suggests, lies in developing "multiplayer harnesses" that extend these superpowers to teams and the entire organization. The YC infrastructure, with its internal tool registry and unified database, serves as a blueprint for this multiplayer future. The registry, which has grown from 20 to over 350 tools, transforms agents from general-purpose assistants into highly specialized, work-relevant tools.

"The first kind of magical moment that I had was we had this agent loop and we had a tool registry, a shared tool registry for YC-specific tools. The first tool that really was an unlock for me was, I think, a tool looking back that you actually built, Jared. It gave these agents the ability to run read-only SQL queries against our database."

The implication here is that traditional security-first approaches, while seemingly prudent, actively hinder the AI's potential. By embracing a "trust-default culture" and broadcasting agent conversations internally, YC fostered a learning environment where employees could observe and adopt AI capabilities organically. This transparency, coupled with broad data access, creates a feedback loop where the system and its users continuously improve, a stark contrast to the siloed, controlled environments common in many organizations.

The Self-Improving Dream Cycle: From Skills to Superintelligence

The most compelling aspect of YC's AI infrastructure is its capacity for self-improvement, moving beyond static prompts and skills to autonomous learning loops. Koomen describes how agents now analyze past conversations to identify areas for improvement and discover missing context. This "self-improving dream cycle," where agents learn from their own interactions and the collective experience of the organization, is the engine of emergent superintelligence.

A prime example is the "two-sentence description" skill. Initially a manual process for YC partners to help founders articulate their companies concisely, this skill was codified, used, and then improved by analyzing meeting transcripts where partners provided feedback. The agent, trained on this collective knowledge, eventually surpassed human proficiency in generating these descriptions. This micro-mechanism, applied across thousands of organizational tasks, is how superintelligence is built.

"This is how super intelligence happens inside organizations. This two-sentence pitch thing sounds like something kind of small, but embedded in it is actually something very powerful... You literally just compose everything that you do and any given thing that any given person can do, you combine that in aggregate and in this particular process, and you have a super organization. It's possible now."

This iterative, data-driven improvement cycle highlights a critical competitive advantage: delayed payoff from immediate discomfort. Building such a system requires significant upfront investment in infrastructure, data unification, and fostering a culture of transparency and trust. Most organizations, focused on immediate ROI and risk-averse, will shy away from this complexity. However, as Koomen implies, those who embrace it will achieve a level of organizational intelligence that rapidly outpaces incumbents. The "horseless carriage" analogy is apt here: AI isn't just an add-on feature to existing software; it's a fundamental shift that redefines how software itself is built and operated, moving control from developers to users and empowering individuals to shape their digital tools.

Key Action Items

Here are actionable takeaways derived from the YC AI infrastructure insights:

• Consolidate Internal Data: Begin the process of unifying disparate data sources into a single, accessible database. Prioritize critical business context.

  • Immediate Action: Identify key data silos and map dependencies.
  • Longer-Term Investment (6-12 months): Architect and implement a data warehouse or lakehouse.

• Develop an Internal Tool Registry: Create a centralized system for cataloging and managing AI-accessible tools and functions specific to your organization.

  • Immediate Action: Start documenting existing internal scripts or APIs that could be exposed to agents.
  • Over the next quarter: Build a basic registry interface and begin integrating a few key tools.

• Foster a Trust-Default and Transparent Culture: Encourage open sharing of agent conversations and development processes. This accelerates learning and adoption.

  • Immediate Action: Pilot broadcasting agent interactions internally, perhaps in a dedicated channel.
  • This pays off in 6-12 months: As employees learn from each other's AI usage, adoption and innovation will accelerate.

• Invest in "Multiplayer" Agent Capabilities: Focus on building AI infrastructure that supports team collaboration, not just individual use.

  • Immediate Action: Explore platforms that allow shared access to agent capabilities and context.
  • This pays off in 12-18 months: Enable teams to leverage AI collectively for complex projects, creating a significant advantage over siloed AI use.

• Embrace Self-Improving AI Loops: Design systems that allow agents to learn from their own interactions and organizational data over time.

  • Immediate Action: Begin collecting and annotating agent conversation logs for potential training data.
  • Requires patience (18-24 months): Develop mechanisms for agents to analyze past performance and suggest or implement improvements to their own skills or prompts.

• Prioritize Agentic Control Over AI Features: Shift focus from adding AI features to building core workflows where AI agents have primary control and deterministic tools serve them.

  • Immediate Action: Re-evaluate upcoming software projects through the lens of agent-first design.
  • This requires a mindset shift (ongoing): Encourage teams to think about how an agent would orchestrate a workflow, rather than how AI can augment a human-driven one.

• Accept the Cost of Token Usage: Budget for the significant token costs associated with extensive AI agent use, viewing it as a strategic investment in future capabilities.

  • Immediate Action: Track current AI usage costs and project potential increases based on planned adoption.
  • This pays off in 12-18 months: Early investment in token usage unlocks the data and learning necessary for advanced AI capabilities, providing a significant head start.