Using Skills with Deep Agents

tl;dr: Anthropic recently introduced something new called agent skills. Think of skills as little folders that contain a SKILL.md file plus any supporting documents or scripts. Agents can discover these folders on their own and load them whenever they need to perform a specific task better. We’ve now added full skills support to deepagents-CLI, making this even easier to use.

The Rise of Generalist Agents

General-purpose agents like Claude Code and Manus have exploded in popularity. With so many people relying on them, you’d expect these agents to use dozens or even hundreds of tools. But strangely enough… they don’t. Claude Code gets by with around a dozen tools, and Manus uses fewer than 20.

At first glance, that feels almost unbelievable — like watching someone build a house with only a hammer and a screwdriver. But the trick is simple: give the agent access to an actual computer. With bash and filesystem access, agents can behave much more like humans. They can poke around, create files, run commands, and solve problems without needing a specialized tool for every tiny task.

This is the philosophy behind deepagents, our open-source agent harness that supports filesystem use and running code. If you’re curious how it works, we walk through everything in our overview video and slides.

Generalist Agents With Skills

So how do these generalist agents manage to handle such a wide range of tasks with so few tools? In our recent webinar, Manus explained the idea beautifully: shift the complexity from tools to the filesystem. Instead of giving an agent hundreds of tightly defined tools, give it scripts, instructions, and resources it can run just like a human would.

Anthropic’s new skills follow the same idea. A skill is simply a folder containing:

• a SKILL.md file with YAML frontmatter + Markdown instructions
• optional supporting resources

Here’s a diagram from Anthropic showing the structure of a SKILL.md file:

(original figure referenced)

Skills bring two big advantages over traditional tools:

1. Better Token Efficiency

Skills load gradually. The agent sees only the YAML frontmatter first and reads the rest only if needed. Tools, on the other hand, usually dump everything into the prompt upfront, which eats up the context window quickly. Anthropic illustrates this clearly in the following diagram:

(original figure referenced)

2. Less Cognitive Overload

Instead of juggling dozens of overlapping tools — which can confuse even the smartest agent — they stick to a few simple atomic tools (like “read a file” or “run a command”). Skills encode the rest.

The result feels cleaner, lighter, and easier for the agent to navigate. Anthropic’s table comparing tools and skills sums this up nicely.

But the exciting part goes beyond efficiency. As Barry Zhang from Anthropic explains, skills push agents toward continuous learning. Agents can create brand-new skills on the fly as they run into unfamiliar challenges. Even better, skills are shareable — one agent can benefit from another’s discoveries. And because skills are composable, agents can mix and match them as needed within a single session.

Deep Agents CLI + Skills

Our deepagent-CLI is an open-source coding assistant designed to behave like Claude Code, with full access to your local machine. Now that we’ve added skill support, you can plug into the growing ecosystem of public skills instantly.

To get started, just create a skills folder and drop in any skill you want from our examples:

mkdir -p ~/.deepagents/agent/skills
cp -r examples/skills/web-research ~/.deepagents/agent/skills/

When you launch deepagents-CLI, it automatically loads any skills found in your agent’s folder. To see what’s available, run:

deepagents skills list

Whenever you ask deepagents to do something related to one of those skills, it will automatically open the corresponding SKILL.md and use it to guide its actions — no extra prompting needed.

If you want a deeper walkthrough, check out our demo video and the full README. We’re excited to see what you build and what new skills emerge as the community grows.