Stop Building Agents. Start Building Skills

Anthropic’s “Skills > Agents” thesis explains why the next wave of AI won’t be a zoo of bespoke bots, but one general agent with a well-stocked toolbox

For the last year, the AI world has behaved like it’s assembling custom robots for every household chore. One agent to write code. Another agent to do research. Another agent to run your calendar. Another agent to “be your COO.” Soon your org chart looks like a sci-fi daycare.

Anthropic researchers Barry Zhang and Mahesh Murag argue that this is the wrong direction. The breakthrough isn’t spawning more agents. It’s equipping a general agent with skills: portable, composable packets of procedural know-how that can be loaded when needed.

https://www.youtube.com/watch?v=CEvIs9y1uog

The problem with agents isn’t intelligence

Models are getting smarter, and “agent scaffolding” is converging. But real work requires context, process, and domain-specific judgment. Zhang’s critique (as reported and summarized from his talk) is that today’s agents often “lack expertise” and miss important real-world context.

Anthropic’s framing is crisp: we’re reaching a point where we can build general-purpose agents that operate in full computing environments (filesystems, code execution, tools). The missing layer is a scalable way to add specialized expertise without rebuilding the agent each time. That missing layer is skills.

Skills are the “applications” layer for agents

In the talk summary, Zhang and Murag lean on a computing analogy: models as the “processor,” agent runtime as the “operating system,” and skills as the app ecosystem that actually encodes useful expertise at scale.

Only a few companies build chips and operating systems. Millions build apps.

Skills are Anthropic’s attempt to make that “millions of apps” layer possible for agentic work.

What a “skill” actually is

Anthropic defines skills as organized folders of instructions, scripts, and resources that an agent can discover and load dynamically to perform better on specific tasks.

The official guide is bluntly practical: a skill is packaged as a simple folder that teaches Claude how to handle a task or workflow, so you don’t have to re-explain your preferences and processes every time.

A typical skill folder contains:

• SKILL.md (required) with YAML frontmatter + instructions

• optional scripts/ for executable code

• optional references/ for docs loaded only when needed

• optional assets/ like templates and brand elements

And crucially: Claude Code treats skills as first-class citizens. Create a SKILL.md, and Claude can use it automatically when relevant or you can invoke it directly with a slash command (e.g., /deploy).

Why “skills” scale better than “agents”

1) Skills use progressive disclosure, so they don’t bloat context

Skills are designed so the model doesn’t guzzle the entire playbook upfront.

At startup, the agent loads only each skill’s name and description. If the skill looks relevant, it then loads the full SKILL.md.

Extra reference files can be pulled in only when needed. Anthropic calls this progressive disclosure and treats it as the core design principle.

This is how you get “unbounded” skill libraries without stuffing every session full of tokens.

2) Skills are composable, so one agent can “multiclass”

Claude can load multiple skills simultaneously, and your skills should play nicely with others (not assume they’re the only rulebook in town). That’s the opposite of “build a new agent for every department.”

3) Skills can include code for reliability and cost

LLMs are great at many things, but some operations are cheaper and more reliable as code. Anthropic explicitly highlights that deterministic code can provide consistency and repeatability, and that skills can bundle scripts the agent can run without loading everything into the context window.

4) Skills can be versioned, tested, shared like software

Murag describes the direction as treating skills like software: testing, versioning, and measuring quality as they evolve.

That’s a big deal. Prompting is often tribal knowledge. Skills turn that tribal knowledge into something reviewable and reusable.

5) Skills aren’t just for engineers

One of the most interesting claims from the talk coverage: Anthropic has seen skills created by people in accounting, legal, recruiting, and other non-technical roles, and Fortune 100 companies using skills as internal “AI playbooks” for best practices. That’s the quiet revolution here: procedural knowledge becomes a file you can ship.

Where MCP fits in (plumbing vs recipes)

The talk summary notes a shift toward skills as MCP becomes a standard for agent connectivity.

Anthropic’s own guide uses a helpful division of labor:

• MCP connects Claude to services and real-time data (tools/connectivity).

• Skills teach Claude how to use those tools effectively (workflows/best practices).

MCP gives the agent hands. Skills teach it technique.

What this looks like in practice inside Claude Code

Claude Code ships with “bundled skills” that are essentially prompt-based playbooks. Examples include:

• /simplify for parallel code review and improvements

• /batch for orchestrating large-scale codebase changes with parallel agents and worktrees

• /debug for troubleshooting a Claude Code session

This is a preview of the world Zhang and Murag are pointing at: one agent, many specialized workflows, invoked as needed. And it’s not just theoretical. Anthropic maintains a public repository of skills and templates, explicitly framing skills as reusable packages that can cover everything from enterprise comms workflows to document tooling.

Adopting a skill-first approach requires a shift in mindset. Instead of aiming for a perfect, all-encompassing agent, focus on creating a portfolio of skills that work together. This approach aligns perfectly with the goal to democratize access to quality sales data and empower entrepreneurs with AI-driven insights and educational resources to achieve significant growth.

Start by evaluating your current processes and identifying where skills can add value. Invest in training your team to think in terms of skills and modular development. Use platforms and tools that support skill creation and integration.

Remember, the future belongs to those who build adaptable, intelligent capabilities—not just agents. By stopping the endless cycle of building agents and starting to build skills, you position your business for sustainable success.

If you want to explore more about how to implement this approach, check out Salesfully, a platform dedicated to helping small businesses and startups boost their sales with AI-driven insights and resources.