Pete Koomen
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
Pete Koomen is the creator of Optimizely, one of the earliest and most prominent A/B testing platforms, and a practitioner-advocate for deploying AI agents at the organizational level. His views are drawn from his talk in the YC AI Playbook series, where he argues for treating AI not as a feature bolted onto software but as the foundational operating layer for entire companies.
AI as OS, Not Copilot
Koomen's central thesis is that companies should avoid treating AI as a mere copilot or feature add-on. Instead, AI should serve as the building layer for everything: "Part of the key thing inside a company is not to just use AI as a co-pilot. This is the thing where you use it as the building layer for everything". He frames the architectural goal as agents wrapping deterministic software tools — not the inverse: "it's going to look a lot more like the agent wrapping software deterministic tools rather than deterministic software wrapping an AI". Corollary to this, he argues that the best AI software is minimal — adding the least code necessary to let the model's capabilities emerge: "The best AI software that I've used tend to be very small".
Database Access as the Key Unlock
Koomen identifies broad, unrestricted agent access to a single shared organizational database — rather than narrowly scoped tools — as the pivotal unlock for practical AI adoption inside companies: "the first tool that really was an unlock for me was a tool that gave these agents the ability to run read-only SQL queries against our database". He reports a Jevons Paradox-like effect: lowering the friction for data queries doesn't just substitute for existing queries — it dramatically expands the volume and ambition of questions employees ask: "it dramatically increased the number of questions that we would ask and dramatically increased the scale and complexity of the questions that we would dare to ask".
He extends this to a structural argument about organizational data architecture: consolidating all context into a single schema gives agents the same advantage that a monorepo gives coding agents: "watching our agents operating on our single database that has everything in one schema tells me that there's a lot of value in getting all of the context into one place".
Tool Registries and the Multiplayer Harness
Koomen identifies a shared tool registry as the primitive that transforms general-purpose agents into useful organizational infrastructure: "The tool registry is what turns these agents into something that's useful at work". The registry provides agents with shared, curated capabilities that can be deployed consistently across the organization.
He flags the multiplayer harness — the infrastructure for enabling agent superpowers at a team or organizational level, rather than just for individual users — as an unsolved problem: "one of the big problems that I don't think has been solved well yet by anybody is the multiplayer harness". This is one of the few areas where he acknowledges an open frontier rather than advocating a specific solution.
Self-Improvement and the Dream Cycle
Koomen describes a "dream cycle" mechanism in which a general agent runs nightly, reviews all employee-agent conversations from the day, and extracts improvements — better context, new skills, corrected failures — to apply autonomously: "we have this general agent that every night will go and read through all of the agent conversations that employees have had and look for things it could have done better". This represents an organizational-level feedback loop for agent self-improvement without human curation of each improvement.
Transparency, Social Learning, and Access Control
Koomen advocates making all agent conversations publicly visible to every employee by default. He argues this simultaneously accelerates organizational learning — employees can see how colleagues use the agent effectively — and creates a social deterrent against misuse: "by defaulting to public broadcast for these conversations, you kind of institute a bit of a social control on what people can do with it". This design choice eliminates the need for explicit policy enforcement by leveraging peer visibility.
Onboarding and Knowledge Democratization
Koomen argues that well-built agent infrastructure compresses the ramp-up time for new employees, effectively giving them immediate access to the organization's accumulated knowledge and best practices: "you could have a new employee joining and maybe would have taken them 6 months to ramp up, but with this it's sort of like they automatically get a lot of the context from the company". This positions the agent infrastructure as a form of institutional memory that is actively accessible rather than passively documented.