The Agent Loop
Every answer is produced by the agent loop — a repeating cycle. In plain terms:
- Take your message (and recall relevant memories, describe any
attachments).
- Assemble the instructions — prompt slots, the tool index,
active skills, data sources.
- Ask the model what to do next.
- If it wants tools, run them (often several at once), feed the results back, and
loop again.
- When it's done, finalise the answer.
Along the way, optional steps can fire: interrupt checks (so a new message can steer
mid-run), permission/guardrail checks (so risky actions pause for your confirmation),
delegation (handing off to another agent), skill loading, and memory saving.
What's powerful is that most of these steps are individually switchable per agent.
A handful of core steps can never be turned off, but around a dozen optional ones can
be enabled, disabled, or made conditional — edited visually in the **Agent Loop
diagram** on the agent card, where disabled steps show muted and struck-through. This
is how an automation agent can, say, skip confirmation prompts while an interactive
one keeps them.
Related: How Agents Work: The Mental Model, Delegation & Hand-off.