Human-in-the-loop
Automation that pauses for human judgment at defined points — approving an outbound email, releasing a payment, accepting a diff. The design question is which decisions warrant a human, not whether automation is safe.
The question is which decisions, not whether
Framing this as a choice between automation and human control produces bad systems in both directions. The useful question is narrower: which specific actions, at which thresholds, warrant a person. Sending external email, moving money, deleting data, deploying code and anything legally binding are the usual candidates.
A gate is only as good as its context
An approval request that says "approve this?" trains people to click yes. A good gate presents what the agent read, what it concluded, what it proposes to do and what happens if it is wrong — enough for a real decision in seconds. Reviewer fatigue is the failure mode that quietly converts oversight into theatre.
Gates should narrow with evidence
Start with approval on everything consequential, then use the accumulated receipts to identify the categories where the agent has been consistently right and raise the threshold there. Oversight that never relaxes is a permanent tax; oversight that relaxes on evidence is a system that earns trust.
In AGNT
AGNT approval gates park a run mid-flight until a person signs off, with the full context attached.
Human-in-the-loop — common questions
Does human review defeat the point of automation?
No — the agent still does the reading, reasoning and drafting, which is most of the work. A human spends seconds approving rather than minutes producing.
Where should approval gates go first?
In front of anything irreversible or externally visible: outbound messages, payments, deletions, deployments.
How do AGNT approval gates work?
The run pauses mid-flight with full context attached and resumes exactly where it stopped once a person signs off.