Skip to main content
With human-in-the-loop confirmations for write actions in the Glean web app, write actions now:
  • Pause before they run, instead of executing immediately.
  • Wait for approval or cancellation before continuing.
This experience:
  • Applies by default to write actions in Agents that run in the Glean web app.
  • Is avaialable only in interactive web app sessions.
  • Excludes read-only actions such as search or lookup.
  • Leaves Slack and Microsoft Teams behavior unchanged.
You can still opt specific steps to Run without user confirmation when the workflow is safe to auto-run in the web app. A write action creates or updates data in another system. Examples include:
  • Creating or updating records in Jira, Salesforce, or other apps.
  • Creating documents or spreadsheets.
  • Sending or drafting emails or messages.
  • Pushing code or opening pull requests with actions like Code Writer.
Human-in-the-loop confirmations do not currently apply to plan and execute steps. Confirmations only appear for direct write actions in interactive web app sessions.

​
Experiencing write confirmations

When a write step starts in the Glean web app, the experience happens in two phases:
  1. Pause with allow/deny The panel shows:
    • Target app, for example, Jira, Salesforce, Google Drive.
    • The specific record or resource when known, for example, issue key, opportunity, document, and so on.
    • A short description of what the agent plans to do.
    From this screen, you can:
    • Allow: let the agent proceed with the write step.
    • Cancel: stop the write action and return to the conversation.
  2. Editable preview (for supported actions) For write actions that support inline editing in the Glean web app, clicking Allow opens an editable preview where you can review and update fields or content before the change is saved. For example, you can:
    • Edit fields on a Salesforce record update.
    • Tweak the text of a Jira comment or description before posting.
    Some actions do not currently have an inline preview. This includes Google and Microsoft actions that create documents, emails, or spreadsheets. For these, Allow runs the action directly, and you can edit the result in the target app or through Canvas experiences that build on the created content.

​
Multiple writes in one run

One of the most powerful capabilities of human-in-the-loop confirmations is the ability to execute complex, multi-step workflows without sacrificing granular control. This ensures that agents can perform tasks while you maintain oversight of every database change or file creation. When a single run requires multiple updates, each write action is presented as an individual confirmation step. This allows for action chaining, where the output of one step becomes the input for the next:
  • Dynamic dependencies: You can approve the creation of a new record and immediately use its generated Record ID or Document URL in the subsequent step of the same run.
  • Granular oversight: Even in a long chain of dependencies, you retain the authority to approve, skip, or stop individual changes.
  • Workflow continuity: The agent pauses for your verification at each critical junction, ensuring that errors in early steps don’t cascade through the rest of the sequence.
This architecture provides the speed of autonomous agents for data handling and the security of human judgment for data integrity. You can automate deep, multi-app processes that were previously too risky to run without constant manual intervention.

​
Example: Project update workflow

Imagine an agent that:
  1. Summarizes recent activity.
  2. Drafts a status update in a Google Doc.
  3. Posts the doc link in a Jira epic.
With confirmations enabled in the Glean web app:
  1. The agent drafts the Google Doc and pauses at the Create Google Doc step.
  2. You review the planned write. For supported actions, you can edit the title or body, then click Allow.
  3. The agent uses the new doc URL to prepare the Jira comment.
  4. You review the Jira update and approve the second write.

​
Controls for agent creators

​
Default behavior for write steps

For interactive agents in the Glean web app, write steps:
  • Pause before every write.
  • Show a confirmation panel with the target app and planned change.
  • Pass write outputs to later steps (IDs, URLs, status fields).

​
Run without user confirmation

Use Run without user confirmation only for safe, repeatable steps in interactive agents:
  • Open the write step in the Agent Builder.
  • Enable Run without user confirmation.
  • Test inputs and parameters in a controlled environment.
When enabled for an interactive agent, the action runs immediately in the Glean web app without showing a confirmation panel. Use this only when you’re confident about the scope and impact of the write step.

​
Scheduled and background agents

Human-in-the-loop confirmations currently apply only to interactive sessions in the Glean web app. For scheduled or background agents:
  • Only a small set of write actions are eligible to run in the background.
  • These eligible write actions run without an interactive confirmation panel.
When Agent Inbox is enabled for an agent, Glean can notify users when a background run is blocked waiting for approval. Users can review and decide whether to proceed from the Inbox.

​
Distinguishing write confirmations from Wait for user input

Wait for user input lets an agent collect structured input during an interactive session. You can use it to ask questions or validate inputs. Write confirmations differ because they:
  • Trigger automatically at each write step.
  • Focus on reviewing and approving a specific change to another system.
You can pair write confirmation with Wait for user input to gather parameters first, then require approval before writing.

​
Best practices

  • Generate content before writing by using a Respond step so users can see and refine what will be written.
  • Reserve inline execution (Run without user confirmation) for trusted, predictable flows in interactive agents.
  • Use scheduled triggers only with pre-approved writes and tight scoping, where you are confident the agent will not attempt unexpected changes.