The Shift to AI-Native Operations

Garry Tan argues that the current AI revolution allows lean teams to achieve output levels previously requiring thousands of employees. The core leverage is not found in model weights—which are accessible to everyone—but in how founders "wire the work." By treating AI agents as a workforce rather than simple autocomplete tools, founders can move from writing individual lines of code to managing a system of encoded capabilities.

Encoding the Organization as Code

To build an AI-native company, every organizational component must be transformed into machine-readable formats:

  • Skill Files: Every task or capability should be documented as a markdown-based skill file. This allows agents to execute repeatable processes reliably.
  • Resolver Tables: These act as the organization's org chart, routing tasks to the appropriate agent or skill file.
  • Performance Reviews: Automated testing and evaluation rules ensure that agents remain in compliance and perform tasks correctly.
  • The Golden Rule: Never perform "one-off" work. If a task is performed once, it should be "skillified" into a reusable file. This prevents the company from "waking up with amnesia" each day and ensures that the organization compounds its intelligence over time.

Context Engineering: The Company Brain

Human working memory is limited to roughly seven items, whereas AI agents can process massive amounts of data. However, the quality of an agent's output depends on "context engineering"—the ability to provide the right information at the right time.

Tan advocates for building a "company brain" (like his project, GBrain), which functions as a library and librarian. The library stores institutional knowledge (emails, meetings, decisions, post-mortems), while the librarian (the retrieval layer) determines which specific information is loaded into the agent's context window for a given task. Without this curation, a company brain becomes a "garbage dump" where stale or contradictory facts lead to confident but incorrect AI outputs. Success requires treating this memory layer with the same rigor as production infrastructure.