Don’t Build Agents. Build Agent Skills for the Web

Over the past year, model intelligence has skyrocketed, and agent tooling has improved significantly. However, there's still a gap: most AI agents don't have the procedural, domain-specific expertise needed to do real work inside real products, especially in web apps, where a small behaviour change turns into user-facing regressions. In this talk, I'll show how leveraging Agent Skills changes how you build AI features on the web: from prompt tweaking to shipping packaged guides, resources and scripts that turn a general-purpose agent into a product-specific one via the Agent Skills protocol (adopted by OpenAI, Anthropic and Cursor). I'll also show a practical way to do this in web apps without changing your current frameworks or LLM, with evals and versioning in place to prevent regressions. ## Outline 1. Title 2. The regressions that break Web AI apps 3. Mega-prompt anti-pattern and why MCPs and tools won’t save you 4. What “Agent Skills” are 5. Architecture (Web ↔ server ↔ agent ↔ VM access) NB: Even though this will be abstracted, I want the audience to understand why it's important. So, I’ll show the abstraction but not shy away from a quick look into its innards 6. Demo: baseline feature a) I’ll start with a simple web app and show why 90% of your users will never be power prompters and how to still have your Agent respond reliably 7. Demo: break it 8. Skill #1: Packaged SOPs a) Show a sample package SOP in an Agent Skill - scripts, validation, etc., all progressively loaded 9. Skill #2: Self learning AI agents with Skills 10. Versioning + tiny evals 1. The current status quo for Evals doesn’t work with Skills. Here, I’ll show how Web engineers can evaluate their AI Agent Skills 11. Checklist - Production checklist for using Agent Skills in a Web Application - regardless of tooling (Vercel AI SDK, langchain) and LLM (works across all LLMs) 12. Closing thesis