Workflow Runner: parallel steps
Steps with no dependency on each other now run concurrently, cutting typical four-step workflow latency by around 40%.
Including the changes that cost us money. A changelog that only lists wins is marketing.
Steps with no dependency on each other now run concurrently, cutting typical four-step workflow latency by around 40%.
Every extracted field now carries its own confidence score, so auto-approval thresholds can be set field by field rather than per document.
Previously a run that failed after the reasoning step was billed at partial cost. It is now free in every failure case, retroactively applied to the prior 30 days.
The agent now surfaces conflicts between sources with both citations, rather than silently preferring one.
Accounts can now set run retention to zero, keeping only billing metadata.
Output now includes questions raised in the meeting that were never answered.
Reduced from $0.006 to $0.004 a run after an efficiency pass. Applied automatically, no action needed.
Agents can now be configured to decline rather than guess below a confidence level you set.