Skip to main content

Human-in-the-Loop - Designing Approval Gates for Agent Workflows

· 3 min read
Calvin Cheng
Calvin Cheng
Shape what gets built and the value it creates.

Autonomous agents can execute high-risk operations: database drops, production deployments, payment API calls. Blocking all such operations defeats automation. Allowing them unsupervised creates unacceptable risk. The solution is human-in-the-loop approval: systematic gates for designated actions, with lightweight feedback that doesn't stall the entire workflow.

The Approval Problem​

Agents need to do things that are irreversible or high-impact. A human would pause before running DROP TABLE users or sending 10,000 emails. An agent, given the right tools, will proceed. The design challenge: where to insert approval, and how to make it fast enough that it doesn't become a bottleneck.

Risk Classification​

Not everything needs approval. Reads, harmless writes, and low-impact actions can run autonomously. The key is classifying operations by risk:

  • High risk: Destructive or irreversible (DROP, DELETE all, production config changes). Require explicit approval.
  • Medium risk: Significant but reversible (bulk updates, external API calls). Maybe approval, maybe thresholds.
  • Low risk: Safe operations. No gates.

Classification should be explicit and configurable. Teams will have different risk tolerance.

Approval UX Patterns​

Channel design: Approvals need to reach the right person quickly. Slack for real-time; email for async; SMS for urgent. The approval request should include: what will happen, why, what data is affected, and whether it's reversible.

Minimal friction: One-click approve/reject. Optional comment for rejections. The goal is a few seconds of attention, not a full review workflow.

Context richness: "Agent wants to DROP table old_users. Reason: cleanup per ticket #1234. Irreversible. Alternative: archive instead." Enough for a human to decide without opening five other tools.

Audit trail: Log who approved what, when, and why. Essential for compliance and debugging.

Design Trade-offs​

Too many gates: Every agent action waits for approval. Throughput collapses. Users get approval fatigue and rubber-stamp.

Too few gates: One bad agent run, one destructive mistake. Trust evaporates.

Wrong placement: Approval for safe operations; no approval for dangerous ones. Inverted risk.

The art is gating only what matters - and making the gates fast enough that they don't undermine the value of automation.

Integration with the Spectrum of Control​

Approval gates sit at the intersection of autonomy and oversight. They let you run agents with high autonomy for most work while retaining human control for the critical slice. The design should feel like a safety net, not a cage: agents can do most things freely; for the rest, a human is in the loop by design.