Where an agent stops and asks
Around 520 BC, Hipparchus of Athens set stone herms along the roads and carved short maxims into them, so anyone walking past could read the law on the messenger's stone. Walk with just intent. Do not deceive a friend.
We run our own company on the same idea. About a dozen (second look) specialist AI agents work behind one front door, and a written permission gate decides what each one may do alone. Money, new prices and first contact with strangers wait for a person.
An agent can build a payment link and check that the page opens. It never presses pay: completing a live payment (STOP) is reserved to the owner, so the agent stops there and reports exactly what is ready.
Every message is reserved before it sends, because a worker that crashes and retries could text the same person twice (CAUTION). We would rather miss one message and see it in the log than send it twice. The loop that answers when an agent gets stuck is public code, with 9 passing tests (GO).
Those colors are Axiōma, the signal our agents use when they write to a person. Red is stop, yellow is caution, green is a check that really passed, and pale gold is a whisper. The signal advises. The gate stays with you. Axiōma is given away: the spec is CC0 and the code is MIT.
# each marked phrase carries its Axiōma signal color and a second cue besides color:
- red, bold, solid line: stop. Do not proceed until it is fixed.
- yellow, wavy line: caution. It works, but it starts something that cuts against your interest.
- green, dashed line: go. A real check passed.
- pale gold, dotted line: whisper. A second look; it stops nothing.
