Enforcement comes before execution
An agent action is an AI interaction like any other. It crosses the same four functions on the way to execution, at the speed agentic systems move.
| ACTIONS ON RECORD | Agent-to-tool calls and agent-to-agent handoffs observed inline, with identity, intent, and destination attached to each |
| PRE-EXECUTION DECISIONS | Allow, block, or hold for human review, applied to the action itself, whatever prompt started it |
| CONTEXT THAT COMPOUNDS | The Security Context Graph™ links agent behavior across sessions, identities, and data sources, so threat and sensitive-data detection grows more accurate as your organization uses AI and connects its data |
| EVIDENCE PER ACTION | Each decision sealed as a metadata-only receipt: what the agent did, which policy judged it, why it was allowed or stopped |
Common questions
What counts as an agent action?
Any operation an agent attempts on the governed path: a model call, a tool call (reading a record, writing a file, sending a message, calling an API), or a handoff to another agent. Agents join the path through the runtime's gateway and its MCP and OAuth tool connections, which carry scoped permissions, so a call runs only within the grant it was given. The runtime treats each action as an AI interaction in its own right, with identity and policy attached before it proceeds.
How do approvals work?
Your policy marks which actions require a person. When an agent reaches one, the action holds before execution: the run pauses, a reviewer sees the action, its context, and the policy that flagged it, and approves or denies it. A held action is the deliberate exception to metadata-only: it is visible to the reviewers you designate while it waits, and once decided, what remains in the trail is the decision. Nothing executes while it waits.
Which actions wait for a person?
Only the ones your policy routes there. Policy separates routine actions, judged inline on the path, from the high-stakes set that waits for a reviewer; everything routine is allowed or blocked at enforcement time, without a queue.
Agents stay within policy
When an agent acts, the action leaves a record: who the agent was acting for, what it tried, which policy judged it, and how the decision landed. If a held action was released, the record names who released it.
