all-agents docs
Changelog

Close the tab and your run keeps going

Sending a message in an interview used to run the entire agent turn inside that one request. Refresh the page, deploy the backend or lose your wifi, and the turn died with it. Your message was gone, and nothing anywhere said so.

Your message and its turn are now recorded before you get an answer. The API replies immediately, the agent works in the background, and every tool call and reply lands in an append-only log as it happens. Reload and you are looking at the same thing you were watching before.

Send two messages in a row and both get answered, each in its own turn, in order. Nothing is dropped for arriving second. A retried send returns the original message rather than posting it twice. A turn that fails records that it failed instead of going quiet.

Runs got the same treatment. A run interrupted by a restart no longer sits at "running" forever: it is picked back up, or reported as failed with a reason.

This entry describes what shipped on the day it shipped. For how the capability behaves now, read /docs/runs.