Agent workflow runs carry on through platform updates

When we ship an update, the servers that run workflows restart one by one. An agent workflow that was partway through, like a scheduled run sending invites or a webhook run updating a CRM, could fail with "This run was interrupted before it could finish", and you had to run it again. Agent workflows and app agents now move to a fresh server and carry on from their last finished step, so the run completes as one run with one result.

How it works

• A clean handover. When a server is told to stop, an agent run finishes the step it is on, saves its files, and goes back to the front of the queue. The next available server picks it up from the saved conversation.
• Finished steps stay finished. The new server continues where the run left off rather than starting over, and the run is billed only for the time it actually ran.
• Every kind of run. A run that is starting, resuming after you answer its question, or continuing after a follow-up message is handed over the same way.
• Recovery from crashes too. If a server stops without warning, the run is picked up once it has missed three heartbeats, usually within a minute, instead of failing several minutes afterwards. A server that only lost touch with the queue keeps its run, so a run never has two servers working on it.
• Code workflows unchanged. A code workflow can't safely restart halfway, so it still gets up to two minutes to finish before an update stops it.