One job, several agents
Split one request across Codex, Claude Code, Cursor Agent, OpenCode and Antigravity, then review and apply one combined change.
Pytxo can give each task in a plan to a different agent CLI. Every task runs in its own isolated copy of your project, tasks that touch the same file run one after another, and you review all the changes together.
Before you start
- At least two agents show Ready in Setup → Agents & permissions. Each one uses its own sign-in.
- OpenCode needs a provider stored with
opencode auth login. Agents start with a clean environment, so API keys exported in your shell do not reach them. - Your project has test or build commands Pytxo can run. Checks are what turn "the agent says it is done" into a result you can trust.
1. Describe the job
Splitting helps when a job has parts that do not depend on each other. You can split it yourself or let an agent do it.
Let an agent split it. Describe the whole job in plain words, then choose Split with OpenAI Codex (or Split with Claude Code). The agent reads your project in its read-only mode and replaces your request with one task per line, each naming the files it owns. Edit any line, or choose Undo to get your words back. Pytxo checks the project before and after and warns you if anything changed while the agent was reading.
Or write the tasks yourself, one per line, naming the files for each part. End a line with | files: to say exactly which files that task owns, including new ones:
Add search and status filtering, with tests | files: src/model.mjs, test/model.test.mjs
Add a dark theme that follows the system setting | files: src/style.css
Translate every interface string to Spanish | files: src/i18n/es.json
In README.md, document the new features and how to run the testsA small fix is usually quicker with one agent.
2. Choose the team
In New work, the Lead agent plans the job and takes the first task. Every other ready agent starts under Agents on this job; untick any you do not want. At most at once sets how many tasks may run in parallel (up to 8). Choose Build plan.
3. Check the plan
Each step lists its tasks, the agent assigned to each and the files it owns. Change a task's agent with its menu or edit its prompt before you run. Tasks that share a file are placed in a later step, so the second agent starts from the first one's result.
4. Watch the fleet
Choose Run. Work opens the fleet board: one card per task with the agent, its recent output and the files it owns.
| Card state | Meaning |
|---|---|
| Working | The agent is running in its own copy |
| Checks passed | The agent finished and Pytxo's checks passed on its task |
| No changes | The agent finished and passed its checks but changed no files |
| Next / Queued | The task waits for a task it depends on |
| Failed | The agent or its checks failed; open the card for its output |
5. Review one change
Choose Review changes. You get one list of changed files, each marked with the agent that wrote it, a line diff for each file, and the checks Pytxo ran on all the changes together.
6. Apply
Apply reviewed changes writes exactly the files you reviewed. If anything in the project changed after review, Apply refuses and writes nothing; choose Refresh review, look again, then Apply. See Review and Apply for interrupted attempts.