# Agent Skills as Codified Domain Expertise
Portable knowledge packages for AI agents. Decision frameworks, edge cases, escalation protocols, all the stuff a 20-year ops veteran carries in their head, formatted so any model can load it instantly.
Evos open-sourced 8 of these across logistics, manufacturing, and energy. Each skill ships with an eval suite of 20-25 real operational scenarios scored against weighted rubrics. Compatible with Claude, OpenAI, Gemini, Cursor, and 26+ other platforms.
This matters because general models can code entire applications but fall apart on a freight claim or energy contract negotiation. The knowledge gap between what AI can reason about and what industry actually requires is the real [[Bottleneck Business]]. Not compute. Not parameters. Domain expertise.
## Why This Connects
**The [[Consultancy-to-Platform Transition]] problem, inverted.** Most deep tech companies get stuck delivering expertise through services. Evos flips it: package the expertise first, distribute it as open source, let the platform emerge around the standard. [[Bespoke Engineering in Industrial AI]] shows that 60-70% of every industrial AI deployment is custom engineering. Skills compress that. Instead of a process engineer spending weeks labeling events and calibrating thresholds, the knowledge is pre-encoded.
**[[Domain Experts as Eval Builders]] in practice.** The eval suites are the real signal. Anyone can write prompts. Evos ships rubric-scored evaluations built from actual operational scenarios. That maps directly to how [[AI Verification]] works: test case libraries built by practitioners, not academics. The rubric is the moat.
**Open source as distribution.** [[Open source is often misunderstood]]. Give away the core, sell the edge. Every platform that loads an Evos skill file is a node in their network. Smart [[AI era Defensibility]] play: win the bailey with distribution velocity, build the motte when you own the standard for how domain knowledge gets packaged for agents.
**Where this sits in the [[AI Agents Stack]].** Skills occupy the tools and knowledge layer. Agents need state management and tool execution, but they also need to know what good looks like in a specific domain. That layer was missing. Evos fills it for industrial ops the same way [[domain specific sense-making]] works: read widely, map ideas, synthesize into reusable frameworks. Except automated and portable.
## The Bigger Pattern
The [[AI Capex Super-Cycle]] flooded the world with inference capacity. Generation is commodity. The constraint moved to: can the model make correct domain-specific decisions? That verification gap, the space between model capability and operational trust, is where the value accrues. [[Where Domain Evals Matter Most]] in high-stakes, regulated, physical-world operations. Exactly where Evos plays.
[[Wright’s Law]] should apply here too. Every additional skill created compresses the next one. The frameworks, eval structures, and rubric patterns become reusable infrastructure. If Evos builds 8, the 80th should take a fraction of the effort.
Links:
- [[Industrial AI MOC]]
- [[Consultancy-to-Platform Transition]]
- [[Domain Experts as Eval Builders]]
- [[Bespoke Engineering in Industrial AI]]
- [[AI Verification]]
- [[Evals]]
- [[Where Domain Evals Matter Most]]
- [[AI Agents Stack]]
- [[AI era Defensibility]]
- [[Open source is often misunderstood]]
- [[Bottleneck Business]]
- [[domain specific sense-making]]
- [[Wright's Law]]
- [[AI Capex Super-Cycle]]
- [[Logistics Optimisation]]
- [[Autonomous Agents]]
---
Tags: #deeptech #kp #systems