My Agents Can Spend Money, So Every Command Has a Tier
A budget caps what agents can spend. It doesn't say which commands they may run. So before agents got paid enrichment in reccy neuro's admin CLI, every paid function was inventoried and every command got a tier.

A budget caps how much your agents can spend. It says nothing about which commands they should be allowed to run.
August's argument was about budgets. On August 18 Fast Company declared that "tokenmaxxing is out, valuemaxxing is in," citing Tesla's cap on employee AI spending, and named the real worry: "An agent running independently at 3 a.m. can't do that, so teams hand it unlimited capacity and walk away."
An agent with production access needs a control on every command.
The call I made
Agents help operate reccy neuro, one of my products, through its admin CLI. In August they got enrichment: refreshing a company's data, sourcing contacts and curating leads. Enrichment calls paid data providers, so the commands that run it spend real money.
The batch refresh across every company is human-only. Single-company enrichment was the open question: one call, one company, real money. I chose to keep it available to agents in production. The pull request's case for that tier is in its open design point: each call is bounded to one company, needs an explicit confirmation flag and is audited.
That leaves one gap: an agent could call the single-company command in a loop. The CLI judges each call on its own. The rule for loops lives in the agent's operator skill: before looping paid calls across many companies, stop and confirm scope.
So the confirmation flag and the human-only check are enforced in code. The loop rule is enforced by instruction. The pull request notes in the same design point that moving the single-company commands to the human-only tier is a one-line change each.
Start with an inventory
Before any command got a tier, every enrichment function was inventoried for two things: whether the backend gates it to admins, and how it spends. One function stood out. The batch refresh defaults to every company, and that's the open-ended bill.
The inventory also cut commands. Four didn't belong in an agent's CLI at all, and they came out before merge, taking the set from 17 commands to 13.
Three tiers
Read. Anything that only reads runs freely.
Bounded write. A paid action limited to a single company runs only with an explicit confirmation flag. The agent has to say, in the command itself, that it means this one.
Danger. The batch refresh across every company, switching features on or off, and global settings are human only. They refuse to run when an agent calls them, so a person has to start them by hand. The batch refresh also rejects a zero or negative company limit, so leaving the limit out is the only way to ask for everything.
That human-only check is a fence against confused plans, not a security boundary against a hostile agent.
Gate first, tiers second
The tiers sat on top of an earlier pull request that added the CLI, its production approval gate and the operator skill. Automated review found ways around the first version of that gate. The pull request stayed unmerged for more than a day, through dozens of automated review comments, until every bypass the review found was closed and re-reviewed. The tiers merged on August 9, on top of the hardened gate.
Where the human should actually sit in an AI workflow was the subject of a May post. In this CLI, it's in front of the commands with the biggest blast radius.
