DeepSeek Harness Automation: Set Up and Verify Tasks
DeepSeek-V4 Team · October 8, 2026 · 7 min read
A reminder that arrives on time is only the first half of an automated task. The second half is the work it asks the model to do. DeepSeek Harness makes that distinction visible through saved delivery records and the linked conversation, but you still need to check both.
In the October 3, 2026 dsh-v0.2.1-alpha.1 release, Automation is built into Web mode. This tutorial follows the version-pinned scheduling guide. It offers a small acceptance procedure for your own installation; it does not report a measured reliability result or promise unattended execution while your machine is off.
Check the mode before writing the reminder
Configure a model provider and use Standard, Creator, or PTC mode. Those presets expose the reminder tools. Minimal mode does not, and delegated subagents are denied access to them. A missing scheduling capability may therefore be a mode issue rather than a wording problem.
Open Automation tasks from the sidebar. The page lets you inspect tasks across conversations without first activating each conversation. Its New action opens a new session with a creation request already in the composer. As with any draft, inspect it before sending.
For the first trial, choose a harmless request with an obvious result. Asking for a short fixed checklist makes delivery and execution easier to inspect than asking the model to search several websites and produce a report. External dependencies can be added after the basic mechanism works.
Start with one delivery you can recognize
Send a request such as:
In two minutes, remind me in this conversation to review the sample report. The reminder instruction should ask you to reply with the three checks: source date, missing rows, and total consistency. Create a one-time reminder, then show me the saved task name and scheduled time.
This asks for a one-time delay, not a repeating task. Open its task details and verify that distinction. Keep the Host running while you wait. When the reminder arrives, inspect the conversation for the requested three checks and open Records to inspect the delivery evidence.
A saved record means the reminder was durably written into the conversation inbox. It does not mean the model finished the requested response. If a record exists but the result is missing, investigate the conversation and provider state before recreating the schedule. Otherwise you risk making several tasks for the same request.
After delivery, a one-time task becomes inactive. Find it with All or Inactive in the task list. Its disappearance from a view that lists active reminders does not mean the task or its history was deleted.
Use a daily rule for a daily local time
A daily wall-clock reminder and a fixed elapsed interval express different intentions. “Every 24 hours” starts from an interval anchor. “Every day at 09:00 in Asia/Shanghai” names a local clock time and a zone. Choose the rule that matches your actual requirement.
For example:
Every day at 09:00 in Asia/Shanghai, remind me in this conversation to inspect the sample report. Name the task Daily report check. Use a daily local-time rule. Do not create a one-time reminder or a fixed 24-hour interval. After creating it, show the saved rule and next scheduled time.
In Rules, check the frequency, time, and saved IANA zone. The next target is displayed in the device's zone, so it can look different from the saved rule if your device uses another zone. Compare the underlying intended time rather than assuming the two visible clock values must match.
Daily schedules in zones with daylight-saving changes need extra care. According to the pinned guide, a nonexistent local time is skipped, and an overlapping time uses the earlier instant once. A fixed interval is elapsed-time scheduling and does not follow those wall-clock rules.
Edit the existing task instead of making a duplicate
Select an active task and open Rules. You can edit its name, instruction, and timing. Edits remain local until Save changes is selected. An accepted edit preserves the task identity, original conversation, and saved delivery records.
If you change the interval, the next interval is measured from the Host's accepted-save time. An equivalent normalized rule does not restart the timer. This difference matters when testing a short interval: changing the instruction alone should not be treated as a fresh schedule.
After saving, reopen or refresh the task and inspect the stored values. If an update cannot be confirmed, inspect the current task before retrying. A stale screen and a failed write are different problems. Inactive tasks are read-only, so do not expect to turn a completed one-shot into a recurring task by editing its inactive card.
Test the Host's restart behavior deliberately
The scheduler runs in the Host. Closing a browser tab is not the same as stopping the Host process. Identify which process or desktop application actually owns your installation before running a restart test.
Use a disposable one-time reminder. Stop the Host before it becomes due, then start it again after that time. The documented behavior is to check stored tasks and deliver an overdue one-shot. For recurring reminders, restart handles only the latest missed occurrence, not every interval that elapsed while the Host was stopped.
Do not use a task with external side effects for this experiment. The documentation does not guarantee exactly-once delivery: a failure after inbox persistence but before the task record is updated can leave a delivered message unrecorded and allow a repeat. A report draft is a better trial than a payment, message send, or irreversible file operation.
Keep an acceptance record with two outcomes
Use a short table for each trial. The distinction between delivery and work completion prevents a green scheduling indicator from hiding an incomplete result.
| Trial | Delivery evidence | Work evidence |
|---|---|---|
| One-time delay | Saved record at the expected occurrence | The requested checklist appears in the conversation |
| Daily rule | Correct frequency, zone, and next target | The next delivered request uses the intended instruction |
| Edit | Same task identity with the new saved values | A later response follows the edited instruction |
| Restart | Expected overdue reminder after Host startup | The linked conversation produces the requested result |
Records are retained within configured history limits. Missing older history cannot be reconstructed simply by looking at the current instruction. Treat the records as delivery evidence for the entries they actually contain, not as a complete audit log of model execution.
Remove the trial when it is finished
Delete a disposable task through its details and confirm the deletion. This stops future delivery and removes the task's saved delivery history. It does not retract a message already queued in the original conversation, so inspect that conversation if work was already due.
The pinned version does not support pausing tasks or creating a new session for every run. Design the workflow around the original conversation and the available editing controls. If you need a continuously operating service with durable execution tracking, establish that separately rather than inferring it from reminder support.
For installation and migration context, read the Harness 0.2.1 update review. Start with a reminder you can observe, then add real work only after you understand how timing, delivery, and completion appear in your installation.