Dify has focused on a very practical direction this time: instead of emphasizing whether workflows can run automatically, it has officially incorporated "when people should intervene" into the system. The official new Human Input node allows the process to be paused at critical links, waiting for someone to review, modify, or press a button to decide the next step, which is more practical for high-risk business scenarios than simply pursuing full automation.
In addition to the human-machine collaboration nodes, 1.13.0 also manually implements the architecture. Officially, workflow streaming and advanced chat execution are added to Celery worker, and events are returned through Redis Pub/Sub, and a 'workflow_based_app_execution' queue is added. For self-hosted and high-throughput teams, this is not an ordinary feature update, but a version change that affects deployment methods and operational configurations.
If we look at this version in the larger competition for AI workflow products, Dify is actually responding to an increasingly clear reality: many companies do not want to completely hand over key decisions to the model, they want "the model runs first, and the person can take over". Whoever can make suspension, review, and resumption of execution into native capabilities will be more likely to eat enterprise-level scenarios.
FAQs
Q: What are the most critical features of Dify 1.13.0 this time?
A: The Human Input node formally puts the manual review into the workflow execution diagram.
Q: Why is this feature important?
A: Because many businesses need to be processed by the model first, but the key nodes still need to retain human judgment.
Q: Does this update only affect product functionality?
A: No, the execution architecture and Celery queue have also changed, so you should pay special attention to self-hosted deployments.
Q: Which teams will focus on this edition?
A: Teams that do enterprise workflows, approval flows, customer service collaboration, and high-risk automation will pay attention.
Q: What trends does this information reflect?
A: AI workflow platforms have begun to make "manual intervention" from patching capabilities to native capabilities.
