Several agents on one task
Let two to four agents try the same task. The first attempt that passes verification wins and the others are cancelled — and what it takes for them to run at once.
A task that changes code can be tried by 2 to 4 agents. In the New task dialog, tick the agents under Try
with several agents; in the API, send attempts:
POST /orgs/:orgId/tasks
{
"projectId": "prj_123",
"title": "Fix the flaky login test",
"prompt": "…",
"attempts": [
{ "agentId": "claude-code" },
{ "agentId": "codex", "providerId": "openai", "modelId": "gpt-5" }
]
}How it works
- Each attempt is a task of its own, pinned to its agent (and provider and model, if given), on its own branch. The task page lists the attempts.
- The first attempt that passes verification wins. The attempts still under way are cancelled, and their workers are told to stop.
- A failed attempt does not end the others.
- Tasks that depend on an attempt depend on that attempt, not on the group.
Running at the same moment
Attempts work in the project's checkout, so a worker runs one attempt of a task at a time. For attempts to run at once you need:
- a worker per attempt, each with the project and the attempt's agent, and
- a project limit (
concurrency.perProject) of at least the number of attempts — the default is 1.
With less, the attempts run one after another, which still gives you the first agent that succeeds.
Cost
Each attempt spends. Spend budgets apply to each attempt on its own, and to the project and organization totals.