Harry Tan on AI-native company: memory is more important than model (Category AI4SDLC)

Reviewed the report by Harry Tan, President and CEO of Y Combinator,Every company should have a BrainAI Engineer World's Fair 2026. The main thesis is that it is not a special model that gives the advantage, but the organization of work around agents and the accumulated context. According to Tan's own assessment, his programming performance has grown to 400 and, with strict amendments, in 8–80 once. This is not a benchmark, but a personal assessment. More importantly, the conditional “2x” and “100x” developers use the same models. The difference is the tie.

Tan suggests looking at an agent system as an organization: - Skill file. - an employee with one duty; - Resolver the organization that directs the task; Storage rules - internal processes; - Evals - check their work (performance review)

Tang proposes to give models and now it is to understand the intention and make a choice. And state, calculations, constraints, and checks can be assigned to deterministic code. Many failures begin when one is replaced by another. But even a good agent is limited to a contextual window, and a company is a library of correspondence, meetings, decisions, and postmortem. Therefore, Tang’s “company brain” is not just a search, but a library plus a librarian choosing the right context for the task.

Tan develops his version in an open project GBrain Memory and search layer for agents with response synthesis, source references, and link graph: Without sources, checking for contradictions, and removing outdated knowledge, such a “brain” would become a good search dump. Memory here is a production infrastructure, not a folder for everything.

Harry's practical advice is not to do one-off work. A successful result should be turned into a repeatable skill and checked through eval. Then the organization accumulates the ability, rather than starting every morning with amnesia.

#AI #AI4SDLC #Agents #Engineering #Architecture #Management