Extend your Agent

Extensions add actions, reusable instructions, external integrations, or durable recall to an Agent. Choose the extension whose ownership and lifetime match the capability you need.

Choose an extension

Extension	Use it for	Owner	Persistence	How the Agent accesses it
Tools	Performing enabled runtime actions	Enabled on an Agent; backing resources keep their own owners	Tool availability follows Agent configuration; effects follow the backing capability	Calls an expanded Tool from an enabled group during a Turn
Skills	Reusing instructions and supporting files	Platform-curated or tenant-owned	Tenant content is stored as immutable versions; the Agent keeps the Skill ID	Receives selection metadata, then activates matching instructions progressively
MCP connections	Calling remote Tools through a credentialed Streamable HTTP connection	Tenant-level connection, attached to an Agent	The connection and credential persist independently of Agent attachments	Discovers the attached connection's remote Tools at Turn startup
Memory	Keeping durable recall across Sessions	Agent-owned, optionally partitioned by `userId`	Persists until updated, deleted, evicted, or the Agent is deleted	Uses Memory Tools or optional automatic context injection

An Agent can combine all four extension types. Configure Tool groups and Memory injection, select Skills, and attach MCP connections before execution; each Turn resolves the applicable configuration and resources.

Tool approvals remain a separate control over sensitive calls, and file Tools still require an attached Sandbox. Coding-assistant integrations belong under Developer Tools, not in the Agent extension model.