AgentScope's new Durable capability covers 5 categories of continuity problems: reconnection after a dropped connection, service restarts, instance replacement, approval waits, and task replay. Our view: the competitive focus of enterprise agents is shifting from "can it complete one inference" to "can it reliably complete tasks that run for hours or even days."
What this is
AgentScope is an enterprise-facing framework for building agents — intelligent programs that call models and tools to complete tasks. The newly added Durable can be understood as the infrastructure layer that lets long-running tasks keep going after a dropped connection, a restart, or an instance swap.
Its core is not saving chat history, but preserving task identity, execution progress, confirmed results, and the conditions for continuing. After a user closes the page, a test can still run; after a service instance is replaced, a new instance can take over; and when an approval arrives, the system can still locate the original task and continue processing it.
The framework also distinguishes two types of instructions: newly added work is queued in order, while supplemental constraints are inserted into the current task. External tools execute asynchronously in an isolated sandbox, and their results are handed back to the original task. That means the agent doesn't have to hold a connection open the entire time, nor depend on any particular running instance.
Industry view
We believe this sits closer to the real threshold for enterprise deployment than simply improving model answer quality. R&D troubleshooting, financial approval, and cross-system operations inherently contain waiting steps; if every dropped connection risks re-execution, the cost isn't only wasted spend — it can also trigger business risks such as duplicate submissions and duplicate charges.
But Durable does not automatically remove those risks. The docs explicitly state that recovery depends on a compatible execution environment and accessible persistent state; if an email has been sent but the receipt is lost, the system still has to verify the original operation. Only when the relevant tools provide query and idempotency capabilities — where repeated execution produces no duplicate side effects — can a task be safely retried.
So what the framework provides is a foundation for reliable execution, not a guarantee of business correctness. Enterprises still have to handle authentication, resource authorization, data access boundaries, and each system's ability to look up prior operations.
Impact on regular people
For enterprise IT: Technical teams will pay more attention to task state, permissions, and exception recovery — not just model selection. Existing agent platforms that lack a persistent execution layer can still see tasks interrupted or operations duplicated during scaling and release.
For individual work: R&D, operations, and customer support staff are more likely to encounter assistants that keep working across sessions. Such an assistant can retain progress and accept mid-run instructions, but whether its results are trustworthy still depends on the data and operation records exposed by the systems it connects to.
For the consumer market: Consumer-facing agents will gradually gain the ability to "keep working in the background after you close the page." Convenience will rise, and users will find it harder to tell exactly which step a task has reached — making in-interface state logs and operation confirmations more important.