Agent audit

Agent audit is for reviewing the permission-decision records of Agents triggering risky operations within the organization, for compliance and security review. When a desktop Agent performs a sensitive action (writing files, running commands, going online, etc.), it goes through a permission decision; this page keeps a complete trail of each decision — what was done, allowed or denied, how risky, and who triggered it.

1 · Overview

Audit records are the “receipts” after the desktop’s permission and sandbox policies take effect. Here you review: which Agents tried to perform high-risk actions, how the system or user ruled on them, and in what execution environment they happened. It doesn’t change client behavior; it makes decisions traceable and reviewable.

2 · Page overview

• Record list — one decision per row, with columns such as tool, executed content, decision, risk, user, and time.
• Filter and pagination — filter by Risky / Normal and page through records.
• Detail drawer — open a single record to view the full context of that action.

〔Figure: the Agent audit record list, with the columns and the “Risky / Normal” filter and pagination highlighted〕

3 · Reading an audit record

A record answers “who, when, wanted to use which tool, did what, and with what result”: a user triggered it, a certain tool at a certain time attempted to execute some executed content, ultimately getting a decision, annotated with the risk level of that action. First use the “Risky / Normal” filter to circle the high-risk actions worth focusing on, then drill down one by one.

4 · Decision, risk, and execution environment

Audit records revolve around several key dimensions:

• Decision result + source tell you “why this was let through / stopped this time” — whether it matched a policy, was manually confirmed by the user, was session auto-allowed, or went to the system default after a timeout.
• Grant scope distinguishes whether this pass applies only to this action or to similar actions across this session.
• Execution environment corresponds to the desktop’s sandbox policy: local execution is the highest risk, while the Windows / macOS sandbox runs in a restricted environment.

〔Figure: the decision / source / risk / execution-environment tags on a record〕

5 · Detail drawer

Open a record and the drawer gives the full context of that action for precise review:

• Parameter summary — an overview of the key parameters of this call.
• Parameter preview (redacted) — a preview of the parameter content, with sensitive information redacted.
• Target resource — the object the action acts on (file, address, etc.).
• Reason — the reason for the decision.

〔Figure: the audit detail drawer, with the parameter summary, redacted preview, target resource, and reason highlighted〕

6 · Important notes

Important notes

• Purpose: audit records are for reviewing high-risk actions of desktop Agents — an after-the-fact tracing tool, not a real-time blocker of actions.
• Pair with Capabilities: when you find a type of Agent frequently triggering risky operations, go back to Capabilities to tighten its distribution or authorization.
• Security tip: focus on the combination of “Allowed + High / Critical risk + local execution” — these records are the most worth reviewing.
• Information redaction: the parameter preview is redacted, but it’s still sensitive audit data — view and retain it within the authorized scope.

Disclaimer: audit data reflects the results of the desktop’s permission decisions, for compliance and security review within the organization; handle it properly per your organization’s policies.

Last updated · July 2026 ｜ Next: Deployment →