Building Agents
MDX
Human-in-the-Loop Patterns for AI Agents
Where approval, review, correction, and escalation create the most value in agent workflows.
Editorial review: clarity, operational relevance, safety boundaries, and source quality.
Approval before consequence
Pause before actions that are expensive, public, destructive, legally meaningful, or difficult to reverse. Show the proposed action, supporting evidence, and expected effect rather than a generic confirmation dialog.
Approval is valuable only when the reviewer can make an informed decision. Present the exact proposed action, changed fields, recipient or destination, evidence, policy exceptions, and expected consequence. Batch low-risk proposals where appropriate; repeated meaningless prompts teach people to approve automatically.
Review after reversible work
For drafts, classifications, or queued changes, asynchronous review can preserve speed. Capture corrections as evaluation data instead of silently discarding them.
Post-action review fits reversible, high-volume work such as labels, drafts, or queue assignments. Sample based on risk and novelty rather than uniformly. Record the original proposal, reviewer edit, and reason code so corrections can improve prompts, tools, policies, or training data.
Escalation is a success path
Agents should recognize missing authority, conflicting evidence, and cases outside their competence. A clean handoff with context is better than a confident but unsupported attempt.
A handoff should preserve work. Include the goal, authenticated identity, facts already checked, sources, attempted actions, current state, and a precise reason for escalation. Measure whether humans must repeat investigation; a high escalation rate can still be healthy if handoffs are accurate and efficient.
Practical example
Reviewing a supplier update
The agent researches a supplier's new bank details and prepares a change showing the old and new values, source documents, identity checks, and risk flags. A finance reviewer approves the exact proposal with a short-lived token. If any field changes afterward, approval is invalidated. The agent then executes the update and rereads the record for confirmation.
Field checklist
Apply it in practice
- Place review at a meaningful commitment point.
- Show evidence and consequences, not a generic prompt.
- Capture edits and reviewer reason codes.
- Make handoffs complete enough to avoid repeated work.
Decision framework
Questions to answer before you build
Human oversight works when people receive the exact proposal, evidence, and consequence at a meaningful commitment point—not when they face frequent generic confirmation prompts.
Review before or after action?
Use pre-approval for consequential work and sampled post-review for reversible, high-volume outputs such as labels and drafts.
What does the reviewer need?
Show changed fields, evidence, policy exceptions, destination, expected effect, and what will remain unchanged.
How will corrections improve the system?
Capture the original proposal, edited result, and reason code, then route findings to prompts, tools, policies, or evaluations.
Common failure signals
Watch for these warning signs
- Approval fatigue caused by interrupting users for low-risk actions.
- Invalidating oversight by changing a proposal after it was approved.
- Escalating without carrying forward facts, sources, and attempted actions.
Selected primary references
Continue with the source material
These sources inform the wider editorial perspective for this topic. They are not presented as line-by-line citations for every statement.