📖What it means

Every agent's chat is a place where work gets DONE, not only discussed. When a Carolverse employee is on the turn, the chat attaches one strict action tool and the model must commit exactly one typed action; deterministic code then checks permission, carries it out once, and stores a receipt. The receipt owns the reply — an agent that claims in prose to have done something without one has that sentence replaced. Side effects fail closed so nothing is half-done; ordinary conversation fails open so a broken boundary can never silence an agent. Until this existed, only the operator's own command line and Carol's chat could make anything happen.

📖Easily confused with

[{"term": "Agent chat as a conversation", "distinction": "A conversation produces words; an action surface produces a receipt. If nothing was written to a record, no action happened \u2014 however confident the reply sounded."}]

📖Nuance

The catalogue of what each agent can do is DATA in the registry, not code: it drives the tool the chat offers, the handbook block in the agent's prompt, and the Handbook app, so a page can never promise an action the chat will refuse. Anything not catalogued is answered with an offer to file an initiative — to perform it once, to make it a permanent action, or both.

📖Also called

- agent chat
- chat action
- action surface

📖Filed under

Category: work.

Service: Agent and Human Resources.