Parallel agents (rolling out)
Split one request into several agents that implement, test, review and integrate together.
One request often holds several jobs: build the change, test it, review it, then put it together. Vixentio can run those as separate agents on one shared work, each in its own copy of the project, with the team watching and steering all of them.
Run a request as several agents
- Write the request in the project's composer (the project must be linked to a team where you're an editor).
- Choose Run as Several Agents….
- Start from the preset Implement + Test + Review, or name 2 to 8 agents yourself, each with a responsibility and a short brief.
- Choose Create Work and Agents. Nothing is created or started before that.
When two or more agents write code, Vixentio adds an integration agent that waits for all of them.
An analysis can also *propose* a split on a Draft. It's an editable suggestion; nothing starts until you approve it.
Dependencies
Agents can wait for others: a Test agent waits for Implement, and the integration waits for every writer. An agent starts only when everything it depends on has finished, and it starts from those agents' changes, so a reviewer reads exactly the change it was waiting for.
- If a dependency fails, is stopped or is lost, the agents waiting on it say so and don't start. Start Again on the dependency picks the chain back up.
- Waiting agents start automatically while the Mac that split the request is running.
Each agent works in its own copy
Every agent but the integration gets its own copy of the project on its host. Agents never write in each other's copies, and nothing reaches your own project checkout except through the integration.
When an agent's change has to reach an agent on another Mac, it travels as a sealed checkpoint: encrypted to the receiving Mac's one-time key, checked when it arrives, and deleted within 24 hours. Teammates never read these packages, and nothing is pushed to GitHub before Publish.
The Agents page
Project › Agents lists every agent across the project's open work: its work, host, status, the last thing it did and what it's waiting on. From a row you can Guide it, Pause or Stop it, start it on your Mac, or move it to another Mac. Role rules are the same as for any run: commenters propose guidance, the driver decides, the host applies.
Early conflict warnings
Vixentio looks for trouble while agents are still working, not only at the end:
- two agents changing the same files;
- two agents targeting the same branch;
- two agents adding database migrations;
- two agents changing dependencies;
- a trial combination of two agents' changes that would conflict.
Warnings name counts and kinds, and only people whose own Mac can see the code are shown file paths. Nothing is merged to resolve a conflict for you.
Integration and review
The integration agent combines the writers' changes one by one. It stops at the first conflict and says where, so a person decides what happens next. The combined result is what the team reviews and, once approved and accepted, what gets published as one pull request. See Reviews and Publish.
What's not there yet
- Agents' copies on a Mac don't run a dev server, so individual agents don't have live previews, and their copies aren't validated on their own.
- Splitting a request and the Agents page are on the Mac only.
- Running parallel agents in cloud workspaces is planned after cloud workspaces themselves.