Human Approval Is a Product Feature, Not a Checkbox
Human Approval Is a Product Feature, Not a Checkbox#
Human approval is usually treated like a safety checkbox.
"We have a human in the loop."
Cool. Where? For what? With what evidence? What happens if they say no?
Approval only works when it is part of the workflow and the UI.
Approval should be specific#
Do not ask:
Approve agent action?
Ask:
Approve posting this incident update to
#status?
Then show:
- Draft message
- Channel
- Source evidence
- Tool calls used
- Confidence
- Risk level
- What changes after approval
The human should not have to guess what they are approving.
What needs approval#
I would require approval before:
- Posting customer-facing messages.
- Issuing refunds or credits.
- Changing account permissions.
- Creating or merging code changes.
- Running deploy actions.
- Deleting, archiving, or mutating records.
- Sending incident conclusions.
- Accessing sensitive customer data outside normal scope.
Read-only lookups can often run automatically. Writes need more care.
Approval should have outcomes#
Approval is not binary enough.
Useful options:
- Approve.
- Reject.
- Edit draft.
- Ask agent for more evidence.
- Reassign to owner.
- Mark as unsafe.
Those outcomes become training data for the workflow. If humans keep asking for more evidence, the Context Scout needs improvement.
Log the decision#
Every approval event should log:
- Who approved.
- What was approved.
- Original agent recommendation.
- Edited final action.
- Sources shown.
- Time to approval.
- Rejection reason.
This makes audits possible and helps improve the agent.
Approval design changes agent behavior#
Once approval exists, the agent can be more useful before the gate.
It can draft. It can collect evidence. It can prepare a ticket. It can explain the risk. It can preview the exact action.
The human handles judgment. The agent handles preparation.
That is the right split.
Build it in Codelit#
Try this:
Build an agent workflow with approval gates for billing changes, deploy actions, GitHub comments, Slack incident updates, and customer-facing messages. Show what evidence the approval UI needs.
Design approval gates in Codelit
Approval is not there to slow the agent down. It is there to let the agent safely do more.
Try it on Codelit
Agent Workflow Builder
Map agents, tools, model routing, approvals, evals, and deployment before wiring connectors
Related articles
Build this agent workflow
Generate a production workflow for Human Approval Is a Product Feature, Not a Checkbox in seconds.
Try it in Codelit →
Comments