Problem
“Human in the loop” sounds responsible, but often it means an ambiguous queue where no one knows when to intervene, what to review, or with what criteria to approve.
The result is poor in both directions: agents too restrained for simple tasks and too free in sensitive decisions.
Human supervision cannot be a permanent red button. It must be a designed capability.
Thesis
Human as Tool is a more operational way to think about supervision.
The human does not appear as a romantic last line of defense. They appear as a specialized tool that the agent can invoke when it detects uncertainty, risk, ambiguity, exception, or high impact.
That requires interface, SLA, context, and criteria.
Framework
A human tool must have five parameters:
- Trigger: when it is invoked.
- Payload: what context the human receives.
- Decision set: what options it can take.
- SLA: how long the agent can wait.
- Learning loop: how the decision improves future rules or evals.
Mini-case: a legal agent prepares contractual responses. It should not request approval for everything. It should escalate only when it detects a change of responsibility, penalties, non‑standard terms, or a strategic client. The human receives a diff, risk, recommendation, and options.
Measurable signal: percentage of human escalations with trigger, decision, and reusable outcome.
Position: supervision without design is friction. Invocable supervision is architecture.
Why it matters now
OpenAI recommends human supervision for sensitive, irreversible, or high‑impact actions. MCP includes elicitation as a pattern to request additional information from the user during an interaction. In agentic systems, the question stops being “human yes or no” and becomes “which human, when, with what context, and to decide what”.
The human becomes a node in the workflow, not a patch.
Anti-example
“Everything is approved by someone before sending.”
It seems safe, but kills speed and does not teach the system. If each approval arrives without taxonomy, the agent does not learn which situations are truly dangerous.
Protocol (3 steps)
- Define escalation triggers. Legal risk, money, sensitive data, irreversibility, low confidence.
- Design minimal payload. Context, agent recommendation, reason for escalation, and options.
- Close learning. Each human decision must feed a rule, eval, or documentation.
| Trigger | Appropriate human | Expected output |
|---|---|---|
| legal risk | legal | approve/reject |
| sensitive data | security | permission |
| key client | account owner | criteria |
| low confidence | domain expert | response |
| irreversible action | manager | authorization |
Related
- Human Escalation Design: when an agent should ask for help and when it should go alone
- Agent Handoffs: frictionless transfers between humans and agents
- Output Verification Layer: the invisible insurance of agents in production
Sources consulted
- OpenAI: A practical guide to building agents
- MCP Elicitation specification
- OpenAI: Evals for AI agents
Next step
Choose a workflow with human approvals. Don’t ask “can we automate it”. Ask which triggers justify a human, what payload is needed, and what decision should be returned.
Translated from the Spanish original with AI assistance and reviewed for accuracy. Read the original in Spanish.