Agent Skills as the New App Store

1 · Anatomy of an Agent Skill

A skill is a folder. At its root sits a SKILL.md file with YAML frontmatter (name, description, allowed-tools, model preferences) and Markdown body containing step-by-step instructions, do/don't examples, and references to scripts. Optional sub-folders hold executable scripts (Python, Bash, Node), templates, and reference documents. The agent loads the skill only when its description matches the user's intent -- preserving context budget. This is the same lazy-loading pattern that made MCP scalable in 2025.

2 · The Four Distribution Channels

Skills now compete for placement on four parallel marketplaces -- each with different economics, governance, and audience. Building once and publishing to all four is the 2026 default playbook. Varun's Daily AI Brief Day 60 · 22 May 2026 · Issue #60

3 · Reference Build -- One Skill, Four Channels

The fastest path to revenue in 2026 is to pick one narrow workflow, write it once as a SKILL.md, then mirror it as an MCP server and register it on A2A -- same artefact, four distribution surfaces. Example: contract-redline, a 4-file skill that takes a Word doc and returns a redlined version using a firm's house

4 · Monetisation Patterns

Skills are unbundling SaaS pricing. Four models now have proof-points in production: per-call (Fluora MCP marketplace + x402 USDC on Base, sub-cent transactions, no signup), subscription (Anthropic Cowork plugin marketplaces, $20-99/user/mo), outcome-based (Agentforce Flex Credits $500 / 100K + 20 credits per action, Claude Managed Agents per task closed), and private offers (enterprise direct, 6-figure deals, 12-mo MSAs). Revenue splits favour the creator: 70-85% to the skill author, 15-30% to the platform -- the App Store curve, but with monthly settlement. Governance trap -- every skill is a privilege-escalation vector. A skill ships allowed-tools in its YAML -- that list IS the trust boundary. Run un-audited community skills inside a sandbox (Day 7), scope OAuth tokens to the skill's named tools only (Day 54 KYA), log every skill load to NIST CAISI WORM audit trail (EU AI Act Annex III evidence), and gate skill marketplaces with reputation scoring before auto-install. The 2025 npm/PyPI supply-chain attacks WILL repeat on

Varun's Daily AI Brief Day 60 · 22 May 2026 · Issue #60

Market Signal

Skills are the protocol layer for the agent economy that MCP was for tool-use. The Stanford AI Index 2026 shows model performance compressing (frontier gap < 3% across Opus 4.7 / GPT-5.5 / Gemini 3.1 Pro / Qwen3.6+), which means the next $50B+ in agentic value is captured by whoever owns distribution + skills ecosystem + governance evidence. Anthropic chose the npm/open-protocol playbook (own the standard, monetise the runtime), Google chose the OS-surface playbook (Gemini Spark Skills inside 3B+ devices), Microsoft chose the enterprise control-plane playbook (Agent 365 as central skill registry). Indie skill-builders win on all three channels with one SKILL.md folder -- the asymmetry of this moment will not last past 2027 when revenue shares formalise and marketplace gatekeeping begins.