The 'Approver' Agent: Adding a Reasoning Layer

An approver agent reads the proposed action, asks the questions a senior on-call would ask, and either approves or kicks it back. The prompt, the eval, and the cost.

The approver's role

The approver does not investigate; it does not act. Its only job is to read the proposal from another agent and decide: approve, reject, or kick back for refinement.

Senior on-call engineers do this naturally. The approver agent codifies the senior on-call's questioning pattern.

Approver is read-only by definition. It cannot modify the proposal; it can only respond to it.

The approver's prompt

"You are reviewing a proposed action by another agent. Ask the questions a senior on-call would ask. Approve only if the proposal answers them all."

Specific questions to require: blast radius, reversibility, evidence quality, alternative considered, escalation path if it fails.

Output: approve / reject / kick-back, with a one-paragraph reason. The reason is logged; future approvers learn from past rejections.

Eval cases

Good proposal: should approve. The approver should not invent objections to a sound proposal.

Bad proposal (over-broad blast radius): should reject. The approver catches what the proposing agent missed.

Borderline proposal: should kick back. "Add the rollback plan and resubmit." The approver demands rigour without rejecting outright.

Cost is real but worth it

Each approver call is one LLM invocation: ~$0.02-0.05. Adds 1-3 seconds to the action latency.

Saves the cost of bad actions. A single bad action averted (a wrong restart, an over-broad change) easily pays for thousands of approver calls.

Justifies the cost only on actions that warrant it. Read-only or trivially-reversible actions do not need approver review.

Approver as a substitute for human review

For routine actions, the approver replaces human review. The senior on-call's time is freed for harder calls.

For irreversible or high-stakes actions, the approver is in addition to human review, not a substitute.

Track which actions get human-only review, approver-only review, both, or neither. Tune as trust grows.