What is the agent inbox?
The agent inbox is where proposed actions from AI agents land when they require human review. Whenever an agent operating in supervised mode — or an autonomous agent whose confidence falls below the threshold — generates a proposal, it appears here for your team to evaluate.
Accessing the inbox
Click Agent Inbox in the left sidebar of your Userorbit workspace. The inbox displays a list of pending proposals, each showing the agent that created it, the type of action, a summary of what it proposes, and a timestamp.
Reviewing a task
- Click on any item in the inbox to open its detail view.
- Read the agent's proposal carefully. For content-related tasks, you will see the full draft text. For categorization or tagging tasks, you will see the suggested labels.
- If the proposal includes generated content, review it for accuracy, tone, and completeness.
Approving a task
If the proposal looks good, click the Approve button. The agent will execute the action — publish the draft, apply the category, send the response, or whatever the proposal entails. Approved tasks move out of the inbox and into the completed section.
For content tasks, you can edit the draft before approving. Click into the content area, make your changes, then approve. The agent uses your edited version.
Rejecting a task
If the proposal is not suitable, click Reject. The task is removed from the inbox and no action is taken. Rejected tasks are logged so you can review rejection patterns over time — this helps you fine-tune agent guidance and confidence thresholds.
Requesting changes
If the proposal is close but needs adjustments, you can send it back to the agent with feedback. Click Request changes and add a note describing what needs to be different. The agent will revise the proposal and resubmit it to the inbox.
Understanding task details
Each inbox item includes:
- Agent name — Which agent generated the proposal.
- Action type — What the agent wants to do (create content, categorize, respond, etc.).
- Confidence score — How certain the agent is about the proposal. Higher scores indicate more certainty.
- Context — The data or event that triggered the proposal.
- Timestamp — When the proposal was generated.
Best practices
- Process the inbox regularly. Pending tasks may become stale if left unreviewed for too long, especially time-sensitive responses.
- Use rejections constructively. If you notice a pattern in what you reject, update the agent's guidance to prevent similar proposals.
- Adjust thresholds. If you are approving nearly everything, consider lowering the confidence threshold so more tasks execute automatically.