Executions & Approvals
Every time a responsibility runs, it creates an execution. You can review what the agent did, what it produced, and approve actions that require your sign-off.
Viewing executions
Navigate to a responsibility to see its executions. Executions are surfaced as chat sessions and are grouped by status in the chat sidebar:
- Working — Currently running.
- Needs Approval — Completed but waiting for a team member to approve an action.
- Done — Finished successfully.
- Skipped — The responsibility’s evaluation criteria determined this trigger didn’t warrant a full run. Skipped executions appear dimmed with a “Skipped” label and are still clickable — open one to see the reasoning for why it was skipped.
Click on any execution to open its chat session.
You can also navigate directly to the filtered execution list from the responsibility overview using the View Executions button, which opens the chat list pre-filtered to that agent and responsibility.
Task list layout
In the task list, the responsibility name is shown as the primary title for each item, with the execution’s generated session title displayed on the second line. This makes it easier to scan which responsibility produced each session at a glance.
Execution titles
Alma automatically generates a concise title for each execution session based on its trigger and completion summary — for example, “Fix login bug escalated” — so you can identify sessions at a glance without opening them.
Execution session header
At the top of each execution chat, the session header displays the status, agent name, and responsibility name as clickable links — selecting either one navigates directly to that agent or responsibility’s settings page. A completion or evaluation summary appears as a footer on the last message once the execution finishes.
Execution details
Each execution shows a step-by-step timeline of what the agent did:
- What data the agent read.
- What reasoning the agent applied.
- Which tools the agent called and what they returned.
- What output the agent produced.
The output of each execution is stored and can be viewed or referenced later.
Approval gates
Approval gates let you keep a human in the loop for sensitive actions.
Tool approvals
When configuring a step within a responsibility, you can require approval before the agent uses a specific tool. For example, you might let an agent read emails freely but require approval before it sends one.
Options:
- Auto-approve — The agent uses the tool without asking.
- Any member — Any team member can approve.
- Specific users — Only designated people can approve.
Output channel approvals
Similarly, you can require approval before a responsibility delivers its output. For example, an agent might draft a Slack message but wait for your approval before posting it.
Reviewing and approving
When an execution needs approval, it appears in the Needs Approval group on the responsibility page. Open the execution to review what the agent wants to do, then approve or reject it.
Approved actions are executed immediately. Rejected actions are discarded and the execution is marked as complete.
Inline tool approvals
Tool approval requests appear directly inside the execution chat, inline with the tool call that triggered them. You don’t need to navigate to a separate inbox — the approval card is shown right where the agent paused, with full context about what it’s asking to do.
The approval card uses plain-language summaries to describe the action clearly — for example, “Send email to alice@example.com” rather than generic “Approve Send Email?” messaging.
The approval card lets you:
- Approve or reject the pending tool call.
- Edit parameters before approving — Structured fields let you modify the action details (such as email recipient, Slack channel, or message content) before the agent proceeds. Slack channel fields show channel names rather than IDs for readability.
- Snooze the request to revisit it later.
- View which permission level is required (any member or specific users).
Inline gate approvals
Human-approval gates — whether on evaluation, step, or acceptance gates — also surface directly inside the execution chat, following the same pattern as tool approvals. When an execution reaches a gate configured for human sign-off, an approval card appears inline in the conversation. You can approve or reject from there without navigating elsewhere.
Approval cards appear as soon as the gate is reached, even before the agent has produced any messages — for example, when an evaluation gate fires at the very start of an execution. The execution progress sidebar also shows a dedicated Approval row (highlighted in amber) when a human approval gate is pending, so you can spot it immediately without scrolling through the chat.
Gate approval cards respect the same audience targeting as tool approvals: gates set to Any member are visible to everyone with access to the session, while gates scoped to Specific users only prompt the designated approvers.