Tools, skills, and subagents

Tools, skills, and subagents solve different problems. A tool performs an action. A skill tells the agent how to approach work. A subagent carries out a delegated task with its own published definition.

Tools perform actions

In Tools, registered tools describe calls to external systems. Their names and descriptions help the model decide when to use them. Their input and answer schemas define the contract; endpoint and authentication settings define how the call reaches the system.

When configuring a tool, give it a clear name, explain when it should be used, set its endpoint and sign-in method, supply the input and answer schemas, and choose its timeout. Follow the schema validation messages. Read and write behavior, approval requirements, and permitted use are part of the tool's setup; a friendly description does not change those rules.

Attach a registered tool in Agent → Capabilities → Tools, or allow authorized discovery when the task benefits from choosing among tools. Use Turn on for me where available to establish the offered environment permission and personal grant. Check the agent's Access section afterward, especially for an API key or another person.

A successful connection does not prove a tool's output is useful. Try it with representative arguments, check the response shape, and consider what the agent should do when the external service fails.

Platform tools provide built-in actions

Open Tools → Platform tools to browse built-in actions, or Agent → Capabilities → Browse tools to see them in the context of that agent. These include capabilities around files, plans, waits, and knowledge.

The catalogue distinguishes tools that are ready from ones requiring more setup or run context. Enabling workspace files, for example, supplies context for file operations. Browsing the catalogue does not independently enable every action or expand permissions.

Skills describe procedures

In Skills, configure a name, stable key, Use when description, Markdown instructions, and optional tags. The description helps an agent decide to load the skill; the instructions supply the procedure, examples, and boundaries once it does.

Attach a registered skill under Capabilities → Skills. Choose Discover when needed for occasional procedures, or Always loaded for guidance required on every run. A skill may refer to a tool, but the run must still be authorized to use that tool.

For the request assistant, a skill could describe how to triage incomplete requests and when to ask for clarification. A lookup tool would separately retrieve the customer's record.

Subagents do delegated work

Publish the specialist first. Under Capabilities → Team, choose Attach an agent and select an available published agent. Give the specialist a focused objective and compatible input/output expectations.

The attachment refers to the selected published content. Publishing a new specialist version does not automatically replace every existing attachment; update and republish the parent deliberately. The parent can discover its bound specialists, but cannot treat any agent in the environment as an authorized team member.

Use delegation when another agent needs its own instructions and capabilities. Use a workflow Agent step when the process must always call a particular agent at a particular point. Those are different forms of coordination.

Discovery and loading

Discovery settings → Where to discover offers two scopes:

• Bindings and other authorized tools and skills: the agent may search beyond its explicit bindings, within authorization.
• Only this agent’s bindings and platform actions: keep discovery within the capabilities assigned to the definition and its built-in actions.

Discover when needed keeps capability definitions out of the initial context until they are needed. Always loaded makes a capability available from the start. Loading preferences apply across tools, skills, and bound subagents; they do not remove access checks.

Tools loaded at a time defaults to 16 and accepts 1–63. The agent can swap active tools as the task changes. This is a context-management limit, separate from the Tool calls limit in Overview, which bounds the number of invocations.

Earlier saved definitions can retain older loading behavior until you publish the new discovery settings. Review the compatibility note shown in the editor when updating an existing agent.