Six human-in-the-loop design patterns for AI agents
This guide outlines six human-in-the-loop patterns for AI agents, ranging from logging only and daily digest reviews to confidence thresholds, draft-then-send, side-effect approval gates, and two-person approval, helping developers effectively manage operational risks.
Added 2026-11-02 · originally published 2026-11-01
Steps
- 1Evaluate the specific action types of your agent and choose the appropriate collaboration pattern for different actions (a single agent can combine multiple patterns).
- 2For read-only queries, use the "Log only" pattern and review audit logs after the fact.
- 3For reversible operations, set up a "Digest review" to let stakeholders review and reverse actions in a daily summary.
- 4Set a "Confidence threshold" so the agent only requests human intervention when scores or rule checks fall below the set line.
- 5For tasks like customer service, use the "Draft then send" pattern where the agent writes emails or tickets for a human to review and send.
- 6For actions with side effects—such as spending money or contacting customers—set up an "Approval gate on side effects" that waits for human confirmation (implementable via LangGraph interruptions or Claude Agent SDK hooks).
- 7Implement "Two-person approval" for refunds or production changes that exceed a set threshold.
Ideas for Chinese platforms
These design patterns translate directly to automated operations on Chinese platforms (such as official account auto-replies, second-hand platform auto-fulfillment and customer service, and community management agents), preventing account bans or financial disputes by using "draft then send" or "side-effect approval" mechanisms.
Please help me design a human-in-the-loop control flow for my AI agent system. My agent's main task is [fill in the agent's specific task, e.g., handling customer service tickets or executing refunds], and I need to implement the following patterns: [fill in the required patterns, e.g., confidence threshold check and side-effect approval gate]. Please provide specific architectural design or code logic recommendations.
Related use cases
Summarized by AI from the original source - refer to the original post for details. Copyright belongs to the original author.