blog · April 3, 2026
The Agent Skills Discovery Problem
Every skills marketplace today is designed for humans. A human browses the catalog, finds a useful skill, installs it, and then — only then — can their AI agent use it. This works fine when humans are the primary users. It breaks down completely when agents are.
What agents actually need
An AI agent cannot open a browser. It cannot search a marketplace. It cannot click install. When an agent encounters a task it hasn't been trained to handle optimally, it needs a way to find relevant guidance at the moment it needs it — automatically, without a human in the loop.
The installation bottleneck
The installation step creates three problems:
Discovery lag
The human has to know the skill exists before the agent can use it. The agent's coverage is limited to the human's awareness.
Coverage gaps
Agents only have access to what the human has pre-installed. Thousands of relevant skills exist; the agent can only use the handful that were manually added.
Staleness
Locally installed skills don't update when the source changes. The agent runs on a frozen copy.
What a registry solves
A registry designed for agents works differently. Instead of pre-installation, the agent queries the registry at runtime — describing what it's trying to do, getting back the right skill for the task. No human in the loop. No pre-configuration per skill.
The agent's coverage expands automatically as new skills are registered. A skill added to the registry today is available to every connected agent tomorrow — without anyone installing anything.
How Invoked works
Invoked connects to your agent as an MCP server. One configuration, done once. After that, your agent can query the full registry on every task — finding skills across domains, across contributors, across use cases — without you having to know they exist in advance.
Skills aren't packages. They're services.
# configure once
see setup instructions →