Production agents need human oversight, not just for safety but for trust, compliance, and continuous improvement. The right level of human involvement depends on the stakes.

Human-in-the-Loop Patterns

Approval gates: Agent pauses and requests human approval before high-stakes actions.

def execute_with_approval(action, risk_level):
    if risk_level == "low":
        return execute(action)  # Auto-execute
    elif risk_level == "medium":
        notify_human(action)  # Notify but proceed
        return execute(action)
    elif risk_level == "high":
        approval = request_human_approval(action)  # Wait for approval
        if approval:
            return execute(action)
        return "Action cancelled by human reviewer"

Escalation triggers: Agent recognizes when it's out of its depth:

• Confidence below threshold
• User expresses frustration
• Task involves regulated domains (medical, legal, financial)
• Error count exceeds limit for current task

Collaborative mode: Human and agent work together:

• Agent proposes actions, human confirms or modifies
• Human provides guidance, agent handles execution details
• Agent drafts, human reviews and approves

Production Deployment Checklist

Infrastructure:

• ☐ Stateless agent service with database-backed state
• ☐ Message queue for async tool execution
• ☐ Timeout handling for every external call
• ☐ Graceful degradation when services are unavailable
• ☐ Auto-scaling based on agent workload

Safety:

• ☐ Maximum iteration limits (prevent runaway agents)
• ☐ Maximum cost per task (prevent budget overruns)
• ☐ Sandboxed code execution (if applicable)
• ☐ PII detection in inputs and outputs
• ☐ Content safety filters on all generated content

Observability:

• ☐ Full trace logging of every agent step
• ☐ Cost tracking per task and per user
• ☐ Latency monitoring (time to complete tasks)
• ☐ Success/failure rate tracking
• ☐ Human escalation rate monitoring

User Experience:

• ☐ Show agent's thinking process (builds trust)
• ☐ Allow users to interrupt and redirect
• ☐ Provide estimated completion time
• ☐ Show progress for multi-step tasks
• ☐ Clear error messages when tasks fail

Evaluation & Improvement

Measure agent performance with:

Continuous improvement loop:

1. Review failed and escalated tasks weekly
2. Identify common failure patterns
3. Improve tools, prompts, or add new capabilities
4. A/B test changes against the previous version
5. Gradually increase autonomy as reliability improves