DeepSeek’s Harness release adds persistent scheduled tasks, background work and support for adding tools during a conversation without invalidating the KV Cache, when the configured model explicitly supports it. It also puts scheduled tasks and time context off by default on Web and Desktop, giving users a say before the agent starts keeping time or running reminders.
DeepSeek Watch analysis
What happened
The v0.1.7-rc.2 release lets users create reminders, inspect run history and set repeats as short as one minute. Tasks survive app restarts, and Desktop tasks keep running when the window closes. Quitting the app brings a warning about what that means for those tasks.
A separate change lets supported models take on newly enabled tools during an ongoing conversation without invalidating the KV Cache, so there is no need to start afresh just because the agent gained a capability. The release also says Auto review can be enabled from the Plugins page; when it denies an operation, users can decide whether to proceed in conversations that allow manual approval.
Why it matters
This is a more useful kind of agent than one that needs a person to restart the conversation every time its toolset changes. Persistent reminders and background work also move the product towards agents that act over time, not just answer one prompt and politely vanish.
There are meaningful controls in the release: scheduled tasks and time context are off by default on Web and Desktop, and quitting the Desktop app warns users about active work. Those defaults matter when an agent can keep doing things after its window closes.
Our read
The most consequential combination is simple: tasks can persist, and the defaults leave activation to the user. That is a sensible starting point for automation. Try the scheduling and background-task controls before relying on them for anything important, and check which models explicitly support dynamic tool additions rather than assuming every setup does.
What to watch
- Whether scheduled tasks and run history work reliably across restarts in practice.
- How users are shown what a background task is doing and when it will run.
- Which configured models support dynamic tool additions without invalidating the KV Cache.
- Whether manual approval remains clear when Auto review denies an operation.
Discussion spark: Should scheduled and background agent tasks stay off by default, or should convenience win once users have clear controls?
Sources and evidence
- v0.1.7-rc.2 (27 September 2026, 05:51 UTC)
not affiliated with or endorsed by DeepSeek