Ant Group announced LingGuang 2.0 on October 27, with rollout in China. According to its release, the agent combines cross-session memory, long-term scheduling, and operation in independent cloud instances. It describes authorized Alipay access and APASS mechanisms separating credentials from the agent and allowing access revocation. These are provider statements, rather than findings from our own audit.
From an answer to a maintained task
Our editorial reading is that the important change is continued work after a conversation ends. An answer to a one-off question can usually be reviewed immediately. A long-running task needs to show what is still pending, what has already finished, and which condition the user has changed since the task was first created.
In our example, someone follows a travel offer but later changes the acceptable date. If the assistant continues comparing against the old conditions, an accurate price can still be irrelevant. A useful test would establish whether the new clarification replaced the old condition inside the task itself. This is an illustrative scenario, rather than an observed LingGuang problem or a report from a customer.
We would also distinguish a notification about an opportunity from an operation already performed. Reporting a change, preparing a proposal, and confirming a booking have different consequences. One generic completed status would conceal that distinction. Evaluating a persistent agent requires visibility into which result was actually achieved, along with the information supporting that status.
Why mentioning MCP is insufficient
The release includes support for MCP and CLI tools. The separate Model Context Protocol specification provides technical context: its OAuth flow concerns HTTP transport, and a server must validate that a token is intended for its own resource. That does not certify the configuration of a particular application or automatically cover every command-line operation.
A compatible interface can simplify connections, but it cannot decide what access a workflow needs. In our proposed pilot, an assistant reading operational notices would start with limited permissions. We would examine what happens when authorization expires and whether an already prepared task stops before its next external action. The test would concern the actual execution boundary, rather than the wording of a permission message alone.
Evaluating long-running work
In our view, an appropriate evaluation includes interruption and recovery, changed conditions, and user cancellation. We would trace whether the system preserves the reason for each decision and whether retrying can repeat an operation that has already completed. These are proposed editorial criteria, not an experiment conducted with the product or a claim of demonstrated defects.
A quiet finish matters too. If nothing significant changes, a persistent assistant should not imply a new result simply because it checked again. Unavailable information should remain distinguishable from a confirmed absence of change. That distinction lets the user understand what the assistant actually knows, instead of mistaking activity for evidence of progress.
We have not connected Alipay or personal accounts, executed LingGuang tasks, or established availability in Bulgaria. What is verified is the announcement of the new version and its stated direction toward continuing assistance. Its value will depend on current task conditions, a traceable status, and reliable completion of the work the user has authorized.
*AI illustration: a conceptual editorial image, not a documentary photograph or an actual product screen.*
*Editorial verification: 2026-10-28T17:14:21.370659+00:00.*
