THE SHORT ANSWER

Give an agent a narrow objective, minimum permissions, explicit stop conditions and observable actions. Require approval for consequential steps, monitor results, route exceptions to a person and keep a named human owner accountable.

Specify the work package

  • Objective and definition of done
  • Allowed tools, data and systems
  • Actions that are prohibited
  • Evidence required before action
  • Approval and escalation points
  • Time, cost and retry limits

Control increases with consequence

An agent drafting an internal summary may need source checking. An agent that messages customers, changes records or spends money needs stronger identity, permissions, approvals, logs and recovery. Use the AI Agents & Automation module for system design details.

Evidence & context: Anthropic · Model Context Protocol

Human-in-the-loop must be usable

A person cannot provide meaningful oversight without context, time, authority and a clear rejection path. Requiring a click while hiding evidence creates ceremonial approval rather than control.

Match the human role—operator, reviewer, decision-maker or affected person—to the actual consequence.

Evidence & context: NIST AI Resource Center

Monitor outcomes and exceptions

Operational record
SignalQuestion
SuccessDid accepted work meet the intended outcome?
ExceptionWhich cases required human handling?
InterventionWhen and why did a person stop the agent?
PermissionDid the agent attempt unnecessary access?
DriftHas the environment or task changed?

Sources & further reading

  1. Building effective agents

    Anthropic. A provider's engineering taxonomy of agents and workflows, not a universal industry definition. We use the conceptual distinction, not its changing product recommendations.

  2. Model Context Protocol authorization

    Model Context Protocol. Official authorization requirements and security considerations. Authentication and authorization remain implementation responsibilities; protocol support is not permission to expose a capability.

  3. AI Risk Management and Human-AI Interaction

    NIST AI Resource Center. Official guidance on different human and AI decision roles and oversight. Appropriate reliance depends on context, consequence and system evidence.

Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.

A correction, a counterexample or an experience worth sharing?

Join the conversation ↗