Notes

Approval Gate

Some actions you are happy for an agent to take a hundred times unattended. Others you want to see before they happen. An approval gate is the mechanism that draws that line: a checkpoint in the loop where the agent pauses, presents what it is about to do, and waits for a person to allow or refuse it.

How a gate is built
The gate belongs to the runtime, not to the prompt. An instruction to "ask before deleting anything" is a request; a gate is a condition evaluated in code before the tool is invoked. Common trigger rules:

  • By tool — anything that sends mail, moves money, writes to production, or deletes.
  • By threshold — a refund under ten pounds proceeds, above it a human decides.
  • By reversibility — saving a draft needs no approval, publishing it does.
  • By novelty — an unfamiliar recipient, an unusual argument, or a step the agent flags itself.

When a gate fires the run has to survive the wait, which is why gates and durable execution travel together: state is checkpointed, the answer may arrive an hour later, and the loop resumes exactly where it stopped.

Getting the placement right
Too few gates and an irreversible mistake ships unseen. Too many and you have rebuilt manual work with extra latency, and reviewers begin approving on reflex — a rubber stamp is worse than no gate at all, because it manufactures an audit trail implying scrutiny nobody applied. Gate on consequence rather than on the agent's uncertainty, and show the reviewer the exact arguments of the pending call instead of a summary of its intent.

An approval gate is a checkpoint in an agent's workflow where execution halts until a human authorises continuation. Gates are placed by policy — before production writes, spending, or external communication — and are declared per tool or risk tier rather than negotiated at run time. The gate is what allows an agent to run unattended across low-risk steps while keeping the consequential ones under human authority.

Think of a junior employee with a company card. They can order stationery on their own, but anything over a few hundred pounds needs a manager's signature first.

An approval gate is that signature step for software working on its own. It runs freely through the safe parts of a job, then stops before anything serious — sending the email, making the payment, deleting the files — and waits for a person to say yes.

The skill is choosing where to put the gates. Too few and mistakes slip through unnoticed. Too many and people start clicking "approve" without reading, which is no protection at all.