The Agentic AI Architecture Rule Everyone Should Know: Fat Skills, Fat Code, Thin Harness

Garry Tan, CEO of Y Combinator, shared what he called "the simplest distillation of what I have learned about agentic engineering this year."

As @garrytan wrote: "Push smart fuzzy operations humans do into markdown skills. Fat skills. Push must-be-perfect deterministic operations into code. Fat code. The harness? Keep it thin."

Fat skills are markdown instruction files for tasks requiring judgment and context: customer emails, interpreting ambiguous requests, contextual decisions. These cannot be encoded as algorithms — they need the model's reasoning.

Fat code is deterministic code for tasks requiring precision: database operations, API calls, calculations. No AI judgment here — only reliable execution.

Thin harness is the orchestration layer between the AI and its tools. Keep it as minimal as possible. No complex routing logic.

A reply crystallized it further: "Do the right job at the right layer. Everything else is architecture astronomy."

First-time agentic builders tend toward two failure modes: putting everything into AI instructions (unpredictability where precision is needed), or encoding everything in code (brittle systems that break on edge cases requiring judgment).

The fat skills / fat code split is a practical heuristic: contextual judgment belongs in skills; precision belongs in code.

For EU development teams building AI automation with GDPR compliance requirements, this architecture has a clean advantage: the deterministic code layer is auditable and loggable, while the AI judgment layer can be scoped to only where determinism is impossible. Compliance audits become significantly cleaner.

Tan also shared a thought about his GBrain project, a semantic knowledge-base layer built on AI:

"The interesting thing about making GBrain is that it feels similar to making a game mod on someone's incredible game engine. And sometimes a sufficiently good Half-life mod becomes Counterstrike."

Today's frontier AI models are the game engine. Applications built on top are the mods. And the best mods become products bigger than what they were built on.

Good agentic engineering has a simple rule: contextual judgment in skills, deterministic logic in code, minimal orchestration in between.

At WebEdge, we build AI automation for Baltic businesses using exactly this kind of pragmatic architecture. Get in touch with our team to learn more.