What Are Skills

Agents can already reason. That doesn't mean they can already do everything.

That's the distinction that makes skills matter. A model can think, explain, summarize, and suggest. A skill gives it a new practical capability — connecting to a service, handling a workflow, controlling a tool, or operating in a domain it couldn't act in before. Reasoning is built in. Capability is often installed.

The Matrix analogy works here: before the right skill, the answer might be "I can help think through that, but I can't actually do it." After the right skill: "I know how to do that now." The tool belt just got bigger.

Think about the gap between understanding what browser research should look like and actually being able to operate a browser. Or between knowing memory should be organized and being able to search it semantically. Skills close that gap between general intelligence and useful execution.

This changes how you diagnose a limitation. If your agent says it can't do something, the answer isn't always "the system can't do that." Sometimes the real answer is "this agent doesn't have that capability installed yet." That's a much better problem to have — because it's often solvable.

Skills come from somewhere, of course. At the broadest level there's a community marketplace where anyone can publish. Useful because it moves faster than any built-in feature set. But broad marketplaces are noisy — some tools are excellent, some are immature, and some aren't worth trusting. That's why there's also a curated discovery layer: a smaller set of tools that have been vetted for quality and safety. You don't need to become a package hunter. You just need to understand that capability is extensible, and that there's a meaningful difference between "anything someone published" and "tools worth serious consideration."

The most important shift: when an agent hits a wall, that wall may be temporary.

Maybe it can think through the problem but needs a skill to act on it. Maybe it can give you a recommendation but needs a skill to gather the information directly. Maybe it understands the workflow but needs a skill to execute it reliably.

Once you see skills this way, the system stops looking like a sealed box with fixed powers. It looks like a teammate with a growing toolkit — not pretending to know everything, but gaining concrete capabilities where the work actually demands them.