Monday Jun 1
STRATEGYAnthropic

Skills, Not Agents

2025NOV
AGENTSSKILLS

Barry Zhang and Mahesh Murag, creators of Anthropic Agent Skills, gave a 15-minute talk at the AI Engineer Code Summit arguing developers should stop building bespoke agents per domain — and instead package domain expertise as composable Skills, markdown files with optional scripts and references that Claude loads only when needed.

Their argument: domain expertise is what the agent is missing. The killer line — "I don't want Mahesh, the 300 IQ mathematical genius, to figure out the 2025 tax code from First Principles. I need Barry, the experienced tax professional." Today's agents are Mahesh; Skills make them Barry.

The architecture: model + agent runtime + MCP servers + Skills library. Skills are loaded only when needed (progressive disclosure — only metadata sits in the model's context by default; the full SKILL.md and its references load on demand). Anyone can write a Skill, anywhere — Git, Google Drive, a zip file. Non-technical people (finance, recruiting, legal) are already writing them inside enterprises. Anthropic deployed Claude to financial services and life sciences with this exact pattern within weeks.

The stack analogy lands the argument: models are processors, agent runtimes are operating systems, Skills are the application layer. Claude's own `skill-creator` skill enables continuous learning — anything Claude writes down is usable by a future version of itself. If you're building on Claude, the question is no longer "what agent should I build?" but "what skills should my team accumulate?"

full brief & sources

⚡ Why this matters

  • Anthropic's clearest articulation yet of how the agent ecosystem should evolve
  • Reframes domain coverage from "build a new agent" to "author a new skill" — an order-of-magnitude shift in effort
  • Surfaces a real distribution channel for vendors: ship a Skill (like Browserbase's Stagehand) and you're integrated with Claude without needing an MCP

🔍 What happened

  • Talk title: "Don't Build Agents, Build Skills Instead"
  • Speakers: Barry Zhang and Mahesh Murag (Anthropic), creators of Agent Skills
  • Venue: AI Engineer Code Summit (AIECS), CODE track lead session
  • 15 minutes; dense; multiple replays recommended
  • Skills = organized folders with a SKILL.md (markdown + front matter) plus optional references and scripts
  • Progressive disclosure: only front-matter metadata is in-context by default; agent loads full SKILL.md when triggered
  • Three skill tiers in the ecosystem: foundational (Anthropic itself), partner (Browserbase, Cadence), enterprise/team (built in-house including by non-technical staff)

💬 Smart takes

  • Barry Zhang (Anthropic): "I don't want Mahesh to figure out the 2025 tax code from First Principles. I need consistent execution from a domain expert."
  • Mahesh Murag (Anthropic): "We're seeing skills built by people that aren't technical — finance, recruiting, accounting, legal. That's early validation that skills make these agents accessible for day-to-day work."
  • Jack Ivers (Crafty CTO / cto4.ai): "15 minutes of solid gold" — frames it as Anthropic's strongest articulation of where the agent ecosystem is going
  • Skeptic — implicit positioning: Anthropic makes Claude AND Skills. The architecture conveniently puts Claude at the application platform's center. Worth checking whether OpenAI / Google describe the world the same way once they ship their own equivalents

🧭 Where this goes

  1. Skills ecosystem grows from thousands at the talk to hundreds of thousands by mid-2026
  2. Non-developer skill authoring becomes a measurable growth vector — finance, legal, HR people become skill builders
  3. Skill marketplaces and sharing tools (Skillport, Cowork) emerge for cross-org distribution
  4. Claude's `skill-creator` enables continuous learning — each session's lessons become reusable skill files for future sessions
  5. The "application layer for AI" framing forces a competing answer from OpenAI and Google — expect parallel offerings within 6 months

🎯 Implication

  • For PMs: Stop building one agent per domain. Build ONE good agent and accumulate a Skills library. Domain coverage becomes additive, not a rebuild
  • For execs: Invest in Skill authorship at the TEAM level, not just engineering. The cleanest leverage comes from non-technical staff packaging their procedures into reusable Skills
  • For founders: The Skills ecosystem is a real distribution channel. Shipping a Skill (paired with an MCP if needed) gives your product a hook into Claude that doesn't require enterprise IT to install your software
·YouTube - Anthropic talk·cto4.ai - Session breakdown·AI Engineer - Code Summit·Anthropic - Claude Agent SDK