Dev.to AI 🤖 Ai 👁 0 📖 3 min read

Approval tiers: when the batch should — and shouldn't — need a human

A reader named Cai left the sharpest comment this project has received, on the Hashnode edition of "Stop approving one by one. Batch it." The argument, in Cai's terms: per-action approval breaks exactly where agents bec

A reader named Cai left the sharpest comment this project has received, on the Hashnode edition of "Stop approving one by one. Batch it."

The argument, in Cai's terms: per-action approval breaks exactly where agents become useful — many small parallel actions. The practical fix is policy-based approval tiers. Routine actions run in batches under an upfront spending or trust limit. Only high-risk outliers escalate. And the batch itself becomes a single signed record, with a countersignature as proof of what ran.

Cai is right, and the comment names the layer the original piece left implicit. Batching was never the destination. Batching is how you earn the data to write a policy.

What a tier actually is

A tier is a pre-decision. You make it once, calmly, with the full context in front of you — instead of forty times, tired, at 10pm, clicking Approve on things you stopped reading at item six.

Three tiers cover most agent work I run:

Tier 1 — run and record. Reversible, cheap, inside a stated limit. File writes in a scratch area, draft creation, read-only lookups. No human in the loop per action. The batch record is the review surface: you read what ran, after it ran, in one place.

Tier 2 — batch and approve. The batch is the unit of approval, not the action. Sends, publishes, and purchases under a limit arrive as one reviewable list: what, to whom, at what cost, with what undo path. One decision covers the set — and the set is small enough to actually read.

Tier 3 — stop and ask, every time. Irreversible, expensive, or outward-facing in a way that cannot be unsent. Money movement above the limit, deletions, anything in another person's name. These never graduate downward just because they went fine last time. "It worked" is not a policy.

Where tiers fail

Two failure modes, both observed, neither hypothetical:

Tier creep. A Tier 2 action that succeeds ten times quietly starts being treated as Tier 1 — by the operator, not by any written rule. The limit that made it safe was the review. Remove the review because the review kept saying yes, and you have rebuilt the rubber stamp one level up. The fix is boring: tier changes are themselves Tier 3 decisions, written down, with a date.

Limits without records. An upfront spending or trust limit only works if the batch record is complete — every action, not the ones the agent chose to summarize. A signed record of a partial batch is worse than no record: it manufactures confidence. This is the Dive problem applied to policy: verify the thing, not the claim about the thing.

The order of operations

If you are starting from per-action approval today, the sequence matters:

  1. Batch first. You cannot write tiers until you can see a week of work as sets instead of a stream of interruptions.
  2. Read the batches. The patterns that belong in Tier 1 announce themselves — same shape, same cost, reversed twice without incident.
  3. Promote deliberately. Write the tier, the limit, and the date. Review the promotions monthly, because the agent's job changes faster than your memory of why a limit was set.

Policy is not the opposite of approval. Policy is approval, moved to the moment when you are actually thinking.

The free sample of Approval-Locked (Chapters 1 + 3, including the Dive) is here: https://veronique2320.gumroad.com/l/sffdfk

The full book: https://veronique2320.gumroad.com/l/approval-locked

Watch the Short this series started from: https://www.youtube.com/shorts/tartUWmzUMw

Field notes as they land: https://operatorfieldnotes.beehiiv.com/subscribe

The index of every pattern so far: https://dev.to/veronicstewart/start-here-agent-operations-that-survive-production-4hm6

If a tier boundary saved your batch — or failed you — repost this for the next operator.

📰 Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.