Garry Tan argues that the value of coding agents depends on how an organisation structures work, not merely which model it rents. Garry Tan compares reusable skill files with clearly defined jobs, routing tables with an organisational chart, and trigger evaluations with checks that the right instructions are actually loaded.
Garry Tan distinguishes model-based judgment from deterministic computation and durable state. The talk argues that a context window should not become the only store for complex operational information: agents need external records and structured systems, while models handle interpretation and less deterministic decisions.
Garry Tan describes a company brain as a knowledge library plus a mechanism for selecting relevant context. Useful memory requires provenance, contradiction handling and curation; otherwise retrieval can confidently surface stale information. The talk also advocates converting lessons from completed tasks into reusable skills so that later work benefits from earlier corrections.
Watch on YouTube
