TokenPad

Safety

Human in the loop

Human in the loop means a person approves or corrects an agent’s action before it takes effect.

It is the only reliable guardrail for irreversible actions. Prompt-level rules can be argued around by an injection; a confirmation step cannot.

The design question is where to place it. Confirming everything destroys the value of automation, so the boundary is usually irreversibility: reads run freely, writes and spends require approval.

In practice

The item worth being strict about is irreversibility. Reversible mistakes are how a system gets better; irreversible ones are incidents. Classify every tool by whether its effect can be undone, and require confirmation on the ones that cannot — delete, send, pay, publish, deploy.

Common questions

Which actions need human approval?

The irreversible ones: delete, send, pay, publish, deploy. Reversible mistakes are how a system improves; irreversible ones are incidents, and the classification is worth making tool by tool.

Does human review slow everything down?

Only if applied uniformly. Gate the consequential minority and let the rest run, then sample the automated path on a schedule — reading what the system actually did finds failures no automated check predicted.

Measure it

  • GuardrailsGuardrails are the rules constraining what a model or agent may do, stated in the prompt and — where it matters — enforced outside the model.
  • Prompt injectionPrompt injection is an attack where text the application inserts into a prompt — a document, a user upload, a tool result — contains instructions the model then follows.
  • AI agentAn AI agent is a loop in which a model receives a request, decides whether to call a tool, receives the result, and repeats until it can answer.

More in Safety