Agent runs resume after restarts without repeating actions

An agent run can take many minutes, and occasionally one of our servers restarts in the middle of it, for example during a deploy. Until July 15, the resumed run could repeat a step it had already taken. Now every tool call with side effects is recorded in a journal before it runs and again when it finishes, so a resumed run picks up exactly where it left off.

How it works

• Journal before and after. A tool call that changes something, like sending an email or updating a CRM record, is written to the journal when it starts and when it completes.
• Rebuilt, not repeated. After a restart, the run rebuilds its unsaved step from the journal instead of executing those calls again.
• Uncertain calls are verified. A call that was in flight when the server stopped is marked as uncertain, and the agent is told to check whether it happened before trying again.
• Reads simply re-run. Read-only calls have no side effects, so they're safely repeated.

Also shipped

• Long conversations with many images stay under the context limit: image reads are stored as pointers to the file instead of inline data, and counted at their real cost.