# Every company should have a Brain Garry Tan, Y Combinator | AI Engineer World's Fair 2026 | 21:08 Source: https://www.youtube.com/watch?v=eBUyTS7SzV4 Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/every-company-should-have-a-brain Published: 2026-07-17 Tags: agents, context-engineering, rag, workflows ## TL;DR - AI productivity depends more on how work is organized around agents than on the model weights themselves. - A company brain should combine a large knowledge library with a librarian that selects the right context for each task. - Teams should turn completed work into reusable skill files instead of repeating one-off tasks. ## Summary Garry Tan argues that AI-native companies can let small teams perform work that previously required much larger organizations. He bases this on his own shift from writing about 14 usable lines of code per day in 2013 to estimating roughly 400 times the output today, while using the same hours or fewer. The difference, he says, comes from wiring work around agents. Skill files define capabilities, resolver tables route tasks, filing rules encode processes, and evaluations test whether the system follows them. Tan extends this model beyond engineering to finance, sales, support, and operations. He also describes a company brain as a library plus a librarian. The library stores institutional knowledge, while the librarian selects the context an agent needs and checks for stale or conflicting information. His practical rule is to skillify every successful task so the organization learns instead of starting each day with amnesia. ## Key ideas ### Agent productivity comes from wiring the work around the model [01:26](https://www.youtube.com/watch?v=eBUyTS7SzV4&t=86s) Tan compares his 2013 output, about 14 usable logical lines of code per day, with his current estimate of roughly 400 times that output. He says the same Claude model is used by people who get very different results, so the difference is not the weights, context window, or API. It is how the work is organized. Even after assuming bloated code, scaffolding, and personal overstatement, he puts the lower estimate at 8 times and the middle estimate at 80 times. At Y Combinator, he says, a quarter of the Winter 2025 batch had codebases that were 95% AI-generated. ### Markdown files can encode the operating parts of an organization [04:04](https://www.youtube.com/watch?v=eBUyTS7SzV4&t=244s) Tan maps agent infrastructure onto company structure. A skill file is an employee with one clearly defined capability. A resolver table is an org chart because it decides which skill handles an incoming task. Filing rules encode internal processes. Trigger evaluations test whether the right skill is loaded and therefore act like performance reviews. In his framing, an AI-native organization is made from these procedures, with engineers maintaining the skills and handling work the skills cannot yet do. The management layer is written into files rather than held only in people's memories. ### Lean AI-native companies are applying the model across the business [05:58](https://www.youtube.com/watch?v=eBUyTS7SzV4&t=358s) Tan gives examples from Y Combinator. Emergence, from Summer 2024, went from public launch to nine figures of annual recurring revenue in eight months, and had 15 people when it crossed $15 million in ARR. Retail, from Winter 2024, reached $60 million with about 40 people. He says these companies encode sales, support, operations, and finance as skills that agents execute. At YC, people outside engineering are also building skill files and cron jobs. He says a finance employee who was not a programmer consolidated about 100 Excel workbooks into one app using OpenClaw and a company brain. ### LLMs should handle judgment while code handles exact computation [08:44](https://www.youtube.com/watch?v=eBUyTS7SzV4&t=524s) Tan separates agent work into latent space and deterministic space. The model is useful for taste, judgment, and interpreting vague human requests. Deterministic code is better for exact storage and computation. He uses a planned Startup School seating experiment as an example. The model can understand who should meet whom, while code should store the positions of 800 people and manage the seating array. That information should not live only in the context window. Tan says the two kinds of computation need to be combined, and many AI engineering bugs happen when work is placed on the wrong side. ### A company brain must choose context from a much larger body of knowledge [10:59](https://www.youtube.com/watch?v=eBUyTS7SzV4&t=659s) Tan contrasts human working memory, which he describes as about seven items, with an agent context of about a million tokens, or roughly a thousand pages. He explains this to his child as three Harry Potter books open at once, with the ability to search and combine them quickly. A company contains far more than three books. Its emails, meetings, decisions, customer conversations, and postmortems form a library. Context engineering decides which three books are open for a particular task. Tan calls a company brain the library plus the librarian, rather than treating retrieval alone as the complete solution. ### Retrieval needs curation, provenance, and contradiction checks [12:54](https://www.youtube.com/watch?v=eBUyTS7SzV4&t=774s) Tan says retrieval is only the primitive. The harder work is deciding what enters the knowledge base, enriching and linking it, separating frequently needed information from cold reference, and resolving disagreements between facts. His own system, GBrain, began as a much smaller collection and grew into a warehouse of about 220,000 pages drawn mostly from his emails, meetings, notes, and experience. He describes it as a second brain that can retrieve earlier conversations and relevant portfolio-company examples before he finishes reading a crisis email. He is also direct about failure modes: an uncured brain becomes a searchable garbage dump, and stale facts can appear with confidence. ### Every successful one-off task should become a reusable skill [15:13](https://www.youtube.com/watch?v=eBUyTS7SzV4&t=913s) Tan's operating rule is to never stop after completing a one-off task. An agent may produce a poor first result, but the user can ask it to fix the work. Once the result is acceptable, the process should be turned into a skill file and loaded into the team's harness. He says that asking for something twice means the first task was not captured properly. An organization that records what it learns becomes more capable each day. One that does not wakes up with amnesia, regardless of model quality. In his view, model quality is rented, while the company brain is owned. ### The proposed company architecture is a thin team, a library, and maintained skills [16:43](https://www.youtube.com/watch?v=eBUyTS7SzV4&t=1003s) Tan tells founders to build an AI-native company from the beginning instead of adding AI to an existing structure. His description includes a thin team, skill files for the work, a company library, a personal AI, and a founder who remains close to the code. He recommends using GBrain if useful, while saying that the concepts work with different tools. He compares OpenClaw with a Ferrari and Codex with a Honda that can accomplish most of the same work. The tools can change, but the ideas of separating computation, using skills as employees, maintaining a librarian, and avoiding one-off work should travel across stacks. ## Notable quotes - "The leverage is not in the weights. It's in how you wire the work." (03:01) - "When you sit down with Claude Code or Codex, you're not writing software, you're hiring, training, and managing a workforce made of markdown." (05:58) - "The question that determines whether your agents are geniuses or goldfish is who decides which three books are open on that desk." (12:30) - "If you have to ask for something twice, you failed." (16:16) - "Abundance is not a policy paper, it is shipped software." (18:42) ## Tools & references mentioned - Y Combinator - Claude - Claude Code - Codex - OpenClaw - GBrain - Startup School - Postgres - TypeScript - Erlang - Elixir ## Who should watch - Founders deciding whether to build an AI-native company rather than add an assistant to an existing workflow. - Engineers designing agent systems that need clear boundaries between model judgment, deterministic code, and stored context. - Teams that repeat operational work and need a practical method for turning completed tasks into reusable procedures.