Lecture Note: Inside YC's AI Playbook
Lecture by: Pete Koomen, General Partner at Y Combinator (ex-Optimizely) Source: Y Combinator Lightcone Podcast Date: May 2026
Executive Summary
Pete Koomen presents Y Combinator's internal playbook for building "superintelligence inside a company" — a framework for turning AI from a productivity tool into the operating system of the entire organisation. The core thesis: make AI the shared organisational brain, not just another tool on the desktop.
The Origin Story
Every good lecture begins with a concrete problem. YC's finance team was the bottleneck — swamped with ad-hoc data requests from partners, portfolio companies, and internal teams. Each query required a human to write a SQL query, run it against the production database, and format the results.
The solution: give AI agents read-only SQL access to the production database.
The result was a Jevons Paradox moment. When the cost of a query dropped to zero, the quantity and complexity of queries exploded. Partners started asking questions they'd never have bothered with before — questions that revealed patterns, anomalies, and opportunities buried in the data.
The Core Thesis
"Building superintelligence inside a company means making AI the operating system — a shared organizational brain."
YC's framework rests on three pillars:
- Agents are the new apps — each specialised agent handles a domain of organisational knowledge
- Tools are the new APIs — agents interact through a shared registry of capabilities
- The organisational brain is emergent — the sum of agents and tools becomes smarter than any individual
Infrastructure Primitives
YC has built four core infrastructure components:
1. Shared Tool Registry
A single catalogue of over 350 tools spanning every team — finance, legal, HR, data science, product, portfolio support. Any agent can request access to any tool. New tools are added continuously as teams identify repetitive tasks that can be automated.
Key insight: Tools must be discoverable, versioned, and documented. A tool registry without good metadata is just a junk drawer.
2. Skills & Meta-Loop
Two key concepts:
-
Skillify — a process that converts successful human-agent interactions into reusable skills. When a partner solves a problem with an agent's help, that interaction can be "skillified" into a new tool available to everyone.
-
DRY & MECE Resolver — a system that continuously scans the skill registry for duplication (DRY = Don't Repeat Yourself) and ensures skills are mutually exclusive and collectively exhaustive (MECE). Overlapping skills create confusion; gaps create failure modes.
3. Self-Improving Dream Cycle
This is where the "superintelligence" label starts to make sense.
Every night, a general agent reviews all conversations that happened during the day. It looks for: - Incomplete or failed interactions - Prompts that led to suboptimal results - Opportunities to combine skills in novel ways
Then it automatically improves prompts and skills based on this analysis. The system gets better at its job without human intervention.
Example given by Koomen: A partner creates a "Two-Sentence Pitch" skill. Other partners generate transcripts from office hours. The overnight agent reads the transcripts, identifies patterns in how successful pitches are described, and improves the original skill. The partner wakes up, tests the improved skill, and says: "This thing is now better than I am... at writing those."
4. Data Denormalisation for Agents
Traditional databases are normalised for storage efficiency — they minimise redundancy. But normalised schemas are terrible for AI agents, which need to retrieve information quickly and with minimal joins.
YC's approach: optimise the data layer for agent retrieval, not human querying. This means:
- Denormalised views that pre-join related data
- Vector embeddings for semantic search
- Cached intermediate representations
- What Koomen calls the "Karpathy Wiki" or "GBrain" concept — a single, unified knowledge store that both humans and agents read from
The Super-Intelligent Organisation in Practice
Koomen illustrates the framework with a concrete example:
- A partner creates a "Two-Sentence Pitch" skill — a prompt that helps startups refine their pitch to two sentences
- During office hours, other partners use the skill and generate transcripts
- The overnight Dream Cycle agent reads the transcripts, identifies patterns in how the best pitches are structured, and improves the underlying prompt
- The next day, the skill is measurably better
- Over weeks, the skill improves beyond what any single human could have written
The outcome: organisational intelligence that exceeds the sum of its parts. The company as a whole becomes better at a task than any individual within it.
Organisational Requirements
Koomen emphasises that this isn't just a technology problem. It requires specific organisational conditions:
Trust-Default Culture
Conversations are globally viewable. There are no private agent interactions. The assumption is that anything you do with an agent benefits from being visible to everyone. This requires psychological safety and a culture that rewards sharing over hoarding knowledge.
Budget Realism
YC spends between $10,000 and $100,000 per year on tokens. That's real money. Koomen's prediction: token costs will drop 10× in the next two years, making this accessible to every startup. But for now, leaders need to budget seriously for agent compute.
Raising the Floor
The biggest impact Koomen observes is not on star performers — they were already productive. The biggest lift comes from raising the floor for new employees. New hires, interns, and junior team members can use the shared brain to become productive in weeks instead of months.
Key Takeaways
| Concept | Insight |
|---|---|
| Jevons Paradox | Lower query costs → more queries → more value |
| Skillify | Turn one-off wins into permanent capabilities |
| Dream Cycle | Let agents improve agents overnight |
| Data Denormalisation | Optimise for agent retrieval, not storage |
| Trust-default | Organisational transparency is a prerequisite |
| Raising the floor | Biggest ROI is on the least experienced team members |
Discussion Questions
- How would a trust-default culture work in your organisation? What would break?
- What are the failure modes of a self-improving Dream Cycle? Could it reinforce bad patterns?
- Is there a limit to how much organisational intelligence can be offloaded to agents? Where's the human-in-the-loop boundary?
- The 350-tool registry at YC — how does a team decide what not to build as a tool?
- If token costs drop 10× in two years, what changes about how we build organisational intelligence?
References
- Source: Inside YC's AI Playbook — Y Combinator Lightcone Podcast
- Related: Jevons Paradox — William Stanley Jevons (1865), The Coal Question
- Related: DRY Principle — Andy Hunt & Dave Thomas, The Pragmatic Programmer
- Related: MECE Principle — Barbara Minto, The Pyramid Principle
- Related: Karpathy's "GBrain" concept — Andrej Karpathy, various talks on AI and software