Performance and long tasks

Image, video, audio, long-document analysis, and multi-step Agent tasks can all take a while. While you wait, watch the status and events instead of resubmitting to see whether the system is alive.

Read the status first

• queued: the task was accepted and is waiting to run.
• running: the task is executing.
• awaiting_approval: an approval needs your attention.
• completed: execution finished — keep verifying the artifacts.
• failed: read the specific error before deciding whether to retry.
• cancelled or expired: the task will not continue; confirm the inputs again and create a new task.

Do not resubmit

Resubmitting the same prompt can create duplicate charges, duplicate media, and several sessions nobody can tell apart. An external Agent should poll events on the original run and read new records through the cursor.

Narrow the task input

• Pass only the source references this task needs.
• Settle the script and shot list before starting video generation.
• Pick the reference images before generating media in bulk.
• Keep large files in Beav; the Agent uses references instead of inline content.

Verify after completion

When a task reports completed, keep checking:

1. The artifacts you need exist.
2. The files or drafts open in Beav.
3. The results are attached to the right workspace and project.
4. The follow-up session saved its session ID.

If the status says completed but the artifacts cannot be read, do not report success. Read the artifact errors or task events first, then decide how to recover.