An agent's job is to call the right tool with the right arguments — check availability, look up a record, send a message — not to generate an answer from memory. That's what makes its actions checkable: a tool call has a name and arguments, not just a paragraph of text to trust.
Which tools an agent can call, what data it can touch, and what it's never allowed to do are configured per workflow, not left to the model's judgment alone. A guardrail is enforced in code before a tool call runs, not just suggested in a prompt.
Anything above a threshold you set — a refund, an external message, a policy exception — pauses for a named human to approve, regardless of how confident the agent is.
Every tool call an agent makes is logged with its inputs and outputs. When something goes wrong, the question is never 'what did the AI do' — it's already answered in the run history.
For the multi-agent runtime that plans, executes, and repairs workflow steps, see AgentOps. For the approval rules and policy guardrails agents run inside, see Governance. For the assistant that handles calendar, reminders, and booking directly, see AI Brain.