taskId | string, required | The task to start the agent on, by the id list_tasks answered; with from, the task of the agent it starts from. (1–36 characters) |
prompt | string | What the agent should do, typed to it as its first message: everything it needs to start on its work, or the developer’s words where they named the job. Every start takes one but a start from another agent, which takes none. (1–64000 characters) |
note | string | Everything else the agent should hear as it starts (what to start with, what to watch out for, what to leave alone), handed to it after the prompt in the same first message; leave it out where there is nothing beyond the prompt. (1–64000 characters) |
title | string | The agent’s title from the start, how the developer sees it named, such as the pull request it will open; leave it out to have it named from what it is told. (1–80 characters) |
branch | string | The branch of the task’s repository to clone the agent’s checkout at and cut its own branch from, such as the head branch of a pull request it builds on; it must already be on GitHub. Leave it out for the repository’s default branch. (1–255 characters) |
model | string | The model to run on, by its catalog id (such as anthropic/claude-opus-5.5), as list_models lists it. Every start names it but one with from, which runs on the model of the agent it starts from unless it names another. (1–200 characters) |
effort | string | The reasoning effort, as the model lists it (such as high); given with the model or not at all. (1–200 characters) |
permissionMode | string, required | How the agent’s tool calls are answered: full_access runs every one unasked, ask asks a member before a file edit or a shell command, and auto leaves each to the harness’s own review, which asks where it cannot decide. A Cursor model takes auto or full_access, and another mode on it is refused. Every start names it, one with from included. (full_access, ask, or auto) |
runsOn | object | Where the agent runs: {“machine”:“vercel”} for the cloud, or {“machine”:“device”,“device”:“<name>”} for one of the member’s own devices (a Mac or a host) by its name, or {“machine”:“device”,“deviceId”:“<id>”} by its id; it must be online, and a name matching none of their devices, or several, is refused with the devices listed. list_machines lists each place with the runsOn to pass. Every start names it but one with from, which runs where the agent it starts from runs unless it names another place, which moves it there. |
from | object | Another agent of your workspace to take over from, on its own task: the new agent is handed its brief and its transcript (with Vespin’s summary of the steps that do not fit) and a copy of its work, runs on its model unless you name one, and hears the note after them; it takes no prompt and no branch. Without messageId it continues the agent, which is stopped; with one it forks the agent at that message, which goes on. |