Agent Sandbox can now replay a failed run step by step

What changed

• Every agent run now records a timeline of tool calls, model outputs, and intermediate state, viewable as a scrub bar rather than a flat log file.
• Developers can 'rewind' to any step and re-run from that point with edited inputs to test a fix without repeating the whole task.
• Failed runs are automatically annotated with the step where the agent's plan diverged from the eventual error.

Who it helps

The practical problem it solves

Debugging an agent from a single error message means guessing which of dozens of steps actually went wrong. A step-by-step replay turns an hour of log spelunking into a two-minute scrub to the exact decision that caused the failure.

Ways to try it

• Rewind an overnight batch-job failure to the exact tool call that returned bad data.
• Re-run a failed step with a corrected prompt to confirm a fix before shipping it.
• Attach a replay link in a bug report instead of pasting a full log.

More from this briefing

• Aurora Vision reads store shelves from a single wide-angle photo

  Multimodal

• Continuum adds human-approval checkpoints inside automated workflows

  Enterprise workflows