100monkeys.ai: for agents, approval is the product

The essay is written from an unusual position: 100monkeys.ai says the piece was drafted in English by its own AI agent and then translated into Japanese by the same agent, and it discloses that AEGIS, Zaru and Carbon Forge are its products. That framing is part of the argument — the company is describing how it wants its own agents governed.

The two 'yes' moments are presented as an ordered hierarchy. Upfront permission belongs in the execution environment because prompt-level rules are only requests that nothing outside the model notices when broken; an architectural rule, by contrast, fails the network call regardless of what the model decides. This extends to defaults: under default-deny, the absence of a 'yes' means 'no', while under default-allow every gap in the policy becomes a permission nobody granted. It adds that agents find those gaps quickly, not out of malice but because they try many things.

Beneath both kinds of approval sits a practical requirement the essay stresses: work must survive. An agent waiting an hour for a human must not lose its place when the process restarts, or the system will either discard the work — pushing people to avoid approving — or start over and risk repeating actions already done. The company's own example is Zaru's email tool, which by default holds a founder's outgoing mail until that specific message is approved in the Zaru web app's Approvals screen.