Don't Build Agents, Build Skills Instead

9:11am - 9:30am | Don’t Build Agents, Build Skills Instead

Speakers: Barry Zhang & Mahesh Murag (both Members of Technical Staff, Anthropic)

Speaker Profiles: Barry Zhang | Mahesh Murag

Bio: Members of Technical Staff, Anthropic

Topic: How Skills are the solution for agents to work reliably in production by packaging procedural knowledge that agents can dynamically load

Notes

• still notice gaps with agents
• not always experiences for how it works
• the reason why we built skills instead of agents
• claude agent SDK does something out the box
• “code is all you need”
  • we don’t need to have multiple agents
• “the agent underneath is more universal that we thought”
  • claude code is general purpose agent
• skills are files and we can use that a way to manage thing
  • can put scripts inside of a skill
• skills are progressively disclosed
  • just the top
  • then when you need more it gets the full file
  • and that can extend to include scripts
• 1000s of skills in the last 5 weeks
• Foundations Skills
  • Document Skills (Anthropic)
  • Scientific Skills (k-dense-ai) scientific research skills
  • Browserbase - automation anything in the browse
  • Notion - notion skills to edit stuff
  • Fortune 100 - org-wide skills
  • Enterprise FinTech for 1000s of SWEs
• trends
  • more complex, production-grade skills
  • complementing MCPs tools
  • non-developers building high-value skills
• Complete picture
  • Agent loop
    • File system
    • MCP Servers
  • Give them a library of skills
• Agent with MCP server and a set of skills
  • Claude for Financial Services
  • Claude for Life Sciences
• How skills evolve in the future
  • Testing and evaluation
• huge value of skills is around sharing and execution
• skill can collect institutional knowledge
• building the skills and sharing them will help make your own agents more capable
• vision of evolving knowledge base
  • especially when claude can make new skills
  • skills makes the concept of memory more tangible
• “We think we’ve converged on the architecture to build agents”

Slides

Slide: 2025-11-21-09-09

Key Point: Demonstrates the rapid growth and scale of the AI Engineer conference series, showing consistent sell-outs and massive audience reach both in-person and online, with YouTube viewership growing from hundreds of thousands to millions.

Literal Content:

• Dark slide titled “Past Event Stats”
• Subtitle: “All events sold out (attendees + expo spaces)”
• Timeline of four events:
  • October 9-10, 2023, San Francisco: AI Engineer - 5,000+ Applicants, 525 Curated Attendees, 29k+ Remote Livestream, 399k+ YouTube Views
  • June 23-25, 2024, San Francisco: World’s Fair - 150+ Speakers, 2,000+ Attendees, 50k+ Remote Livestream, 800k+ YouTube Views
  • February 19-21, 2025, New York: AI Engineer - 5,000+ Applicants, 800 Curated Attendees, 150k+ Remote Livestream, 1.5mm YouTube Views
  • June 3-5, 2025, San Francisco: World’s Fair - 200+ Speakers, 3,300+ Attendees, 150k+ Remote Livestream, 4.5mm YouTube Views

Slide: 2025-11-21-09-10

Key Point: Appears to be contrasting Google’s Gemini 3 announcement against Meta’s Llama 4 release, possibly commenting on marketing approaches or product positioning (the labels “Kino” vs “Slop” suggest a value judgment about the quality or presentation of these announcements).

Literal Content:

• Split screen comparison showing:
  • Left side labeled “Kino”: Google’s announcement page for “Gemini 3” dated Nov 18, 2025, titled “A new era of intelligence with Gemini 3” with author photos and Gemini 3 logo
  • Right side labeled “Slop”: An article about “Llama 4 herd” dated April 5, 2025, showing a diagram of “Llama 4: Leading Multimodal Intelligence” with four variants (Behemoth, Maverick, Scout)

Slide: 2025-11-21-09-11

Key Point: Contrasts evidence-based projections of AI capabilities (based on actual data) versus speculative/overhyped predictions about future AI capabilities - highlighting the difference between realistic expectations and hype.

Literal Content:

• Title: “The same AI chart = Kino vs Slop”
• Two charts side by side:
  • Left chart (Meta logo): Shows “The time-horizon of software engineering tasks different LLMs can complete 50% of the time” with timeline from 2012-2026, showing progression of capabilities
  • Right chart: “AI 2027” showing progression of models (OpenAI, Claude, Google Gemini Ultra) from 2027 onwards
• Bottom labels: “Evidence” (left) and “Fan Fiction” (right)

Slide: 2025-11-21-09-12

Key Point: A critical statement about AI-generated content quality - it takes significantly more human effort, judgment, and taste to filter out low-quality AI outputs (“slop”) than it does to generate them, highlighting the challenge of maintaining quality standards in an AI-abundant world.

Literal Content:

• Text in purple: “The amount of taste needed to fight slop is an order of magnitude bigger than that needed to produce it.”

Slide: 2025-11-21-09-17

Key Point: Chronicles Anthropic’s development timeline in 2025 showing their progressive release of agent-related technologies: from building effective agents, to establishing MCP (Model Context Protocol), to launching Claude Code, and finally the Claude Agent SDK - demonstrating their strategic evolution in the agent space.

Literal Content:

• Title: “Year of agents”
• Timeline from FEB to TODAY showing progression:
  • “How We Build Effective Agents” (Anthropic)
  • “Model Context Protocol” with adoption graph
  • “CLAUDE CODE” logo
  • “Claude Agent SDK” with icons
• Anthropic branding at bottom

Slide: 2025-11-21-09-20

Key Point: Demystifies the concept of “Skills” in their system - they’re simply organized as file system folders containing markdown documentation and Python scripts, making them accessible and easy to understand for developers.

Literal Content:

• Title: “Skills are just folders”
• Shows folder structure:

  anthropic_brand/
  ├─ SKILL.md
  ├─ docs.md
  ├─ slide-decks.md
  └─ apply_template.py

• Anthropic branding at bottom

Slide: 2025-11-21-09-27

Key Point: Illustrates the three key aspects of Skills development lifecycle: systematic evaluation/testing, version control over time, and composability where skills can be combined or used with other tools (like MCP and pandas) to create more complex capabilities.

Literal Content:

• Title: “Exploring how Skills evolve”
• Three columns showing:
  1. Evaluation: “Brand Style Skill” with test results (triggering accuracy 8/10, output quality matched, script execution failed, SKILL.md structure valid)
  2. Versioning: Timeline showing dates 2025-10-31, 2025-11-21, 2025-11-27
  3. Composability: Diagram showing “Branded Decks” composed of “Brand Style Skill” + “PowerPoint Skill”, and separately “PowerPoint Skill” connected to “MCP” and “pandas”
• Anthropic branding at bottom