Keep long-running work moving
Hand complex, long-running tasks to the Agent.
Some complex tasks fit in one sentence but take many turns to finish. Set a persistent goal in the current conversation, and Levin checks the evidence at the end of each turn. If the goal is not yet complete, Levin starts the next turn automatically until the work finishes, pauses, is cleared, or reaches its usage limit.
Define the finish line first
Use regular chat for a short question. Set a goal when the work needs several turns, repeated checks, or time to wait for a background task. The clearer and more verifiable the goal, the more reliable each completion decision becomes.
Step 1Turn on Goal mode
Select the Goal control below the chat composer.
Step 2Write the goal
State the output, constraints, and evidence that will count as complete.
Step 3Check the Goal bar
After sending, review the goal status, current iteration, and next step above the chat.
For example, “Run the ProteinMPNN plugin tests and complete one minimal design task” is easier to verify than “Test the plugin.”
How a goal keeps moving
A conversation can have one goal at a time. At the end of every turn, Levin checks the actual evidence, decides whether the goal is complete, and determines whether another turn should begin.
| State | What Levin records | What you can do |
|---|---|---|
| Running | The current iteration, completed checklist items, tool activity, and output. | Add constraints or step in to change direction. |
| Continuing | The current verification result and next action. | Wait for the next turn. You do not need to keep sending “continue.” |
| Paused | Completed turns, tool records, and current progress. | Answer a question, edit the goal, or resume later. |
| Complete | The completion decision, verification evidence, and final summary. | Review the result and begin the next task. |
Verification relies on evidence such as changed files, command output, and test results. A plan, a to-do list, or time spent running does not count as completion.
Step in at any time
A goal can change after it starts. There is no special syntax to remember. Tell Levin to change direction, pause, or clear the goal, and it will replace, pause, or remove the current objective.
| Interface | Role in the goal | Evidence to review |
|---|---|---|
| Edit goal | Change the direction, scope, constraints, or completion criteria. | The revised objective and current status in the Goal card. |
| Chat workspace | Runs each turn and shows messages, tool activity, and results. | File changes, command output, and test results. |
| Goal summary | Records checklist items, progress, and the next step for each iteration. | What this turn completed and what remains. |
The goal defines how much work counts as complete. The execution mode decides which actions require approval.
Review before resuming
When a goal stops, read the reason before continuing. A reliable recovery keeps the original objective intact.
- Read the current phase. Confirm which step was active and what it was expected to produce.
- Check the dependency. Look for a failed task, missing approval, unavailable tool, or changed project file.
- Make the smallest correction. Adjust the context or phase input without restarting unrelated work.
- Verify the new evidence. Continue only when the result meets the phase's completion condition.
Monitor jobs and keep the execution trail
Long-running jobs can keep working in the background while Levin records their status, tool calls, outputs, and errors. When a job finishes, its returned artifact path stays with the record too. You can see what is still running, what has finished, and where to continue or investigate without reconstructing what happened.
Confirm the goal is complete
Completion is a recorded state, not a sentence in a model response. Before sharing the result, open the final phase and confirm that every required output and check is present.
- Final phase: the last phase is marked complete, not skipped.
- Evidence: the expected file, structure, measurement, log, or report is attached to the phase.
- Checks: validation commands or review notes appear in execution history.
- Boundary: no approval, user decision, or failed background task remains unresolved.