What is human-in-the-loop AI? Keep control of agent actions
Human approval puts a deliberate checkpoint between agent reasoning and a consequential action. The agent can prepare the work, while a person decides whether it should be applied.
The concept
A plain-language definition before the product implementation.
Approval is a runtime boundary, not a polite prompt
Telling a model to ask first is weaker than enforcing a gate in the tool and workflow runtime. A real approval step suspends the run or queues the pending action, preserves the proposed arguments, and resumes only after an authorized decision.
How AI Agent uses it
Review the exact action before execution
AI Agent uses approval steps in workflows and pending actions around gated tool calls. Review the proposed change, approve or reject it, and keep the decision connected to the run that produced it.
Exact proposal
Review the action and its arguments rather than approving a vague summary.
Durable pause
Preserve workflow state while the run waits for a person.
Decision trail
Keep approval or rejection attached to the work and its evidence.
From idea to a working system.
- 01
Prepare
The agent or workflow reaches a gated action with concrete inputs.
- 02
Review
A person inspects the proposal and surrounding evidence.
- 03
Resume or reject
Continue with approved data or stop the proposed action.
Gate consequences, not every lookup
Best for
Write or destructive actions where a person should inspect the exact proposal and its evidence before execution.
Choose another pattern when
Routine read-only retrieval. Requiring approval for harmless lookups adds friction without creating a meaningful safety boundary.
Included in AI Agent
The product capabilities behind the idea.
- Workflow approval steps
- Pending tool actions
- Suspend and resume
- Approve and reject history
Frequently asked questions
Keep exploring
Related product guides
Workflows
A workflow turns a business process into an explicit graph. The model can reason inside an agent step, while the surrounding sequence keeps triggers, tools, branches, approvals, and outputs predictable.
Autopilots
An autopilot is a workflow promoted into recurring operations. It runs a defined agent and runbook on a schedule while keeping customer-facing or destructive actions behind approval.
Chat
Agent chat is the conversational control surface for real work. Ask a question, attach context, choose an agent and skills, then review the evidence, tool activity, and proposed changes in the same thread.
Build the smallest useful version first.
Start with the agent and one real task. Add workflow control, context, and approvals when the work shows you where they matter.