First mission
Give one agent one small task, check the plan, and review the result in Desktop.
Start with one bug fix or one test in a repository you know, and one agent you already use: Codex, Claude Code, Cursor Agent, OpenCode or Antigravity. Once that works, try one job with several agents.
1. Describe the result
Open Work → New work. Confirm the project and the agent, then enter a concrete request. For example, replace these paths with files that exist in your project:
Add input validation in src/api.ts.
Reject an empty name and cover it in test/api.test.ts.Name the expected behavior and the files the agent may change. If Pytxo cannot determine the file scope, make the request more specific.
2. Check the plan
Choose Build plan, then inspect:
- Agent: the agent you chose is installed and signed in.
- Files: the allowed paths cover the task without including unrelated work.
- Checks: the test or build commands are appropriate for the requested change.
- Blockers: resolve anything shown before starting.
If no checks are detected, add a known working command under Additional verification commands and build the plan again. Do the same after changing the agent or configuration.
Local planning does not call a model provider by default. Running an agent uses its own account and may incur provider charges. A missing cost estimate does not mean a run is free.
3. Run the task
Choose Run. Follow the task on Work. Select a worker to inspect its status or open its output. A worker saying it is finished is not a passing test result or a confirmed Apply.
4. Review the result
Choose Review changes from Work, or open the run from History.
Read the prepared file changes and the checks recorded for the combined result. Checks run in a separate snapshot; dependencies, build outputs, or Git history may be excluded. A check that needs an excluded file must be addressed before treating the result as verified.
Choose Apply only when you accept the changes. If the project has changed since review, Apply refuses; refresh the review and inspect the new result. Read Review and Apply for blocked or interrupted attempts.
Try another task
Use New work for the next request. Saved requests can provide a starting point, but always confirm the repository, agent, paths, and new plan. Earlier checks and approvals do not carry over.
Prefer the terminal?
The CLI reference covers pytxo mission, previewing a plan, and running explicit tasks. Since 1.3 the CLI can also review and apply a run: pytxo review <run> and pytxo apply <run>. For a scheduling example without a model account, try the collision demo.