Skills
Skills are how the agent gains capabilities. A skill bundles instructions with the tools those instructions need. When a skill is active, its instructions are woven into the system hint and its tools are available for the turn.
Skill activation
| Kind | When active | Examples |
|---|---|---|
| Always-active | Present in every chat from the start | Knowledge, intents, awareness, web search |
| On-demand | Activated by the agent when it needs it: it’s hinted about relevant skills and can search the catalog for more | Email, calendar, reminders, self-modification |
Some on-demand skills also declare a required capability rather than (or in addition to) built-in tools. Email is one: it has no tools of its own and only appears once an integration providing the email capability is connected.
Tool composition
The active skill set is additive. Each active skill contributes its instruction to the system hint and its tools to the turn; the agent can call the union of all active skills’ tools.
A skill usually owns a coherent group of tools but can reuse another skill’s when its framing calls for it: planning reuses the calendar skill’s calendar tools and the reminders skill’s reminder tools. Because the tool set is a union, a shared tool is offered once no matter how many skills require it. Newly activated tools become available on the next model round, including within the current turn.
A skill needs all the capabilities it requires — its instructions describe tools each one provides, so it only advertises when every one of them is connected. Capabilities that improve it without being necessary are optional: planning takes calendar and reminders as optional, working with whatever is connected and never asking.
Flows
A skill’s instruction can lay out a whole multi-step process, but models tend to read it all and try to do everything at once. Flows fix this by drip-feeding the instructions one stage at a time.
A skill declares named flows, each an ordered list of stages. The agent moves through a flow with two tools:
start_flow: begins a flow, activating stage one.advance_flow: closes the current stage and returns the next, until the flow ends.
Only the current stage’s instructions are shown, so the agent works one step at a time and stays on track even if the conversation drifts.