Skills & Tools
What an agent can actually do.
A reasoning model can produce text. An agent can act — send the email, write the record, query the ledger, read the document. Skills are what turn one into the other.
A skill is a single action the agent can take. The set of skills attached to an agent is its complete vocabulary of actions, and the planner chooses from exactly that set. Nothing else is available to it, whatever the instructions say.
The list is the planner's universe. A skill that isn't attached cannot be used. A skill that is attached but whose description doesn't match the requests users actually send won't get picked. Auditing that list is the highest-value thing you can do when an agent isn't behaving as expected — see Skills & Capabilities.
Where skills come from
Four sources, all landing in the same list.
Platform-supplied. Every agent starts with a set attached — the basic verbs: show something to the user, reason about a topic, search the web, query attached knowledge, list session files. They arrive automatically and can be individually disabled. Turn off what the agent doesn't need; an unused skill is one more thing the planner might wrongly choose.
The Skill Library. First-party skills you attach deliberately — API calls, document handling, PII scrubbing, classification, mapping. See the Skill Library reference for each one.
Connectors. Connecting an integration registers its skills into the agent automatically. Two catalogs:
Pre-built connectors to recognized SaaS products — Stripe, Resend, Twilio
The action involves a product with a logo and a stable customer-facing API
Model Context Protocol servers — market data, scraping, search, knowledge retrieval
The action involves a data API or a generic tool
The catalogs overlap. When both have what you want, pick the one that fits your authentication situation — you may already have OAuth set up for one, or an API key ready for the other.
Custom skills. Anything your team authors, when neither catalog covers it. The API Call skill is the general-purpose escape hatch for services that aren't in either catalog.
Skills versus protocols
A skill is an action. A protocol is a way of reaching things. MCP is a protocol: connecting one MCP server can register several skills at once. That's why MCP appears both here and in Protocols — it's the mechanism by which a class of skills arrives.
Why fewer is often better
It is tempting to attach everything, on the theory that a more capable agent is a better agent. It usually isn't.
Every attached skill is a choice the planner has to make correctly. Skills the agent doesn't need widen the space of things it can do wrong — including things that are hard to reverse.
So the pre-attached set is worth reviewing rather than accepting. Open Skills & Capabilities on a new agent and read the list: anything there that your agent's job doesn't require is a capability you can disable, and each one you remove is a category of unintended action the agent can no longer take.
Attach what the agent's job requires. Remove what it doesn't.
Where this is configured
Skills & Capabilities — the full list, and the planner's universe
Adding a Skill to the Agent — the flow for attaching one
Apps & Services — SaaS connectors, per agent
MCP Tools — MCP servers, per agent
Skill Library — reference for each first-party skill
Last updated