When an Agent Needs a Human Checkpoint

Agents fail quietly when nobody defined what they may not decide.

A checkpoint is not a vibe. It is a hard stop in the path where a named role must approve before the next tool call or send.

At Portage (fictional), an ops agent could draft refund emails and mark cases resolved. It closed a high-value account dispute without finance review. The draft was polite. The refund was wrong. The missing gate cost more than the automation saved that quarter.

Decisions that stay human-owned

• Money movement above a set threshold.
• Customer commitments on price, legal terms, or SLAs.
• Access changes that widen who can see sensitive data.
• Any action that deletes or overwrites system-of-record fields.

How to write the gate into the workflow

• Name the role, not a team alias.
• State what the agent may prepare before the gate.
• State what it may not do until approval lands.
• Log the approve or reject with a timestamp.

Portage added a finance checkpoint on refunds over a fixed amount. Volume of drafts stayed high. Wrong closes stopped. Becuase the path finally matched the risk.

Leave-behind

• Which agent actions can spend money or make promises?
• Who must click before those run?
• Where is that gate visible in the tool path?