[Feature]: Export and Replay Agent Workflows

[nayar-900]

Which destkop app does this feature request relate to?

• Select a project 👇
• Agent-TARS (cli, server, agent, tool etc.)
• UI-TARS Desktop

What problem does this feature solve?

Many users perform repetitive desktop workflows using UI-TARS, such as:

• Filling forms
• Navigating websites
• Downloading reports
• Managing files and folders
• Testing application interfaces

Currently, when a successful task is completed, the workflow cannot be easily saved and reused. Users must repeat the same instructions every time or rely on the model to rediscover the workflow.

This becomes inefficient for recurring tasks and business processes that require consistent execution.

What does the proposed features look like?

Add a workflow recording and replay feature that allows users to save successful agent executions as reusable templates.

Possible functionality:

Save Workflow

After a task completes successfully:

• Save the sequence of actions
• Save prompts and parameters
• Store screenshots or checkpoints (optional)

Replay Workflow

Allow users to:

• Re-run a saved workflow with one click
• Modify variables before execution
• Schedule recurring runs

Export / Import

Support exporting workflows as:

• JSON
• YAML

This would allow sharing workflows across teams and devices.

Example Use Cases

• Daily report downloads
• Automated website monitoring
• Repetitive data entry tasks
• QA testing procedures
• Browser-based business workflows

Suggested UI

New sidebar section:

Workflows
├── Saved Workflows
├── Recent Runs
├── Import Workflow
└── Export Workflow

This would make UI-TARS more useful for automation-heavy users and reduce repeated prompting for routine tasks.

[HarperZ9]

Useful feature request. For replay to hold up in business workflows, I would separate the reusable template from the run receipt:

• Template: goal, allowed apps/sites, variable slots, preconditions, credential boundaries, steps with selectors/actions, expected checkpoints, and stop conditions.
• Run receipt: template version, variable values, model/tool versions, checkpoint hashes, observed window/DOM state, action outputs, timings/retries, failures, and any human approvals.
• Replay verifier: compare the current run to stored checkpoints and return MATCH / DRIFT / UNVERIFIABLE instead of silently adapting through changed state.
• Export/import: include schema_version, a redaction map, and an explicit rule that secrets/cookies/session material never leave the local profile.

That gives users the practical "one click rerun" path without turning a task that worked once into brittle hidden automation. I am building Project Telos around the same receipt-loop problem, so I would be happy to test a JSON/YAML schema against messy workflows if useful: https://harperz9.github.io/field-guide.html