Where should humans remain? — designing approval gates for AI agents

Every organization that deploys AI agents hits the same dilemma first. If a human verifies everything the agents produce, the gains of automation disappear. If no one verifies anything, an incident is a matter of time. The answer isn’t how much to verify — it’s where to place verification.

From in-the-loop to on-the-loop

The principle the industry is converging on is the shift from human-in-the-loop to human-on-the-loop. Instead of approving every step, the human sets objectives, constraints, and boundaries, and supervises the orchestration as a whole. The human’s role changes from reviewer to boundary designer — deciding which actions an agent may simply take, and which must stop and ask.

What belongs in an approval request

2026 agent-governance practice has converged on a concrete shape for approval requests. A request arriving at a gate should carry five things — the proposed action and its parameters, the agent’s reasoning trace, the estimated impact and side effects, a rollback procedure if one exists, and an expiry time on the approval itself. Without those five, the human isn’t approving; they’re rubber-stamping, and the gate becomes theater.

Three criteria for placing a gate

Where to put gates can be judged on three axes. Is it hard to reverse? (A deploy can be rolled back; a sent email cannot.) Does it touch the outside? (An internal draft and a customer-facing reply are different classes.) Does it require judgment? (Is there a trade-off that doesn’t reduce to rules?) If all three are low, automate. If any one is high, gate it.

Regulation is moving the same direction

This is no longer just best practice. Singapore IMDA’s Model AI Governance Framework for Agentic AI (January 2026) is the first comprehensive framework for autonomous agents — requiring a verifiable digital identity per agent and an audit trail of which agent acted under whose authorization. The EU AI Act requires human-oversight mechanisms and log retention for agents in high-risk domains by August 2, 2026. Gates and audit trails are becoming table stakes, not differentiators.

No observability, no gates

Gates only work under one precondition — you can see what your agents are doing. Agent observability practice is settling into four layers — metrics, traces, logs, and evals — with an eval-to-guardrail lifecycle where pre-deployment evaluation criteria become production guardrails.

How we apply this

Notique runs execution in three columns — automated, human approval gate, still human. Release decisions, content publishing, and customer replies sit at gates; contracts and regulatory accountability sit in the human column. We don’t claim this placement is universally right — we publish it in How we run and move things when they turn out to be wrong.