The loop

Propose, perform, return, propose again: the cycle that makes an assistant an agent.

A chat writes. An agent writes and acts. The difference is a loop: the model proposes an action, the ordinary software around it performs that action — search the web, open or save a file, send a message, run a sum — and brings back the result, and the model proposes the next. The model writes; the program acts. That is the whole of it, and it is worth knowing because it explains both the reach and the limits.

It is load-bearing because it says exactly where capability comes from. An agent can do only what the program around it can do; a tool it does not have is a thing it cannot do, however fluently it discusses doing it. It also says where the risk sits: each turn of the loop is an action in the world, taken on a reading of text that may have come from anywhere. Hence the rule that governs every turn — what an agent reads is information, never instruction — and the guard that can stop an action before it happens.

A Player+ uses the loop knowingly. Ask your agent what it can do besides talk, and check the answer against what actually happens. Keep the things that send, post, share or spend on the asks-first list; those are the turns of the loop that cannot be taken back. Where an agent works while you are not watching, it works only within what you authorized in advance, in your own words.

The loop is also why a set-up problem shows as a stall: the model proposed something the program could not perform, or a guard blocked it, and nothing came back.

Also called: — Stands on: Language model · Set-up · Agent Opens onto: Safety check · Authorized · The maker · Skill · Venue card In play: beyond Sources: Player+ Modules, How an AI Agent Works, Becoming a Player+, Jam Conventions · The DNA of Heaven, Part X. Open: none found.