Parallel subtasks and Best-of-N

​

Overview

• run_parallel: independent subtasks, one per computer, run at the same time. Each result comes back in input order.
• Best-of-N (run_subtask with best_of: N): for a high-value subtask, N rollouts of the same subtask run at once on N sandbox computers. A judge picks one by the rollouts’ step narratives. This is Agent S3’s Behavior Best-of-N design (Apache-2.0).

​

run_parallel

{
  "member": "run_parallel",
  "input": {
    "subtasks": [
      { "computer": "cmp_build_box", "goal": "Open the build page and start the nightly build", "success": [{ "text": "Build queued" }] },
      { "computer": "cmp_report_box", "goal": "Export the weekly report as PDF", "success": [{ "text": "Export complete" }] }
    ],
    "max_parallel": 2,
    "budget_ms": 120000,
    "max_cost_usd": 0.25
  }
}

• Each subtask takes run_subtask’s fields plus computer, the computer id it runs on. Without computer it runs on the computer the call was made on.
• Each subtask goes through its own computer’s gates: target, control lease, policy, a durable audit row, then the subtask loop. Its steps then go through lease, policy and audit as usual.
• One input owner per computer. A computer listed twice is refused before anything runs (409 computer_conflict; this-device and its own id count as one). A computer that another subtask loop is already driving, from this call or another, comes back refused with error: "computer_busy". On a cloud computer, a lease held by another agent run also refuses.
• Approval: one approval covers the whole call when any subtask targets a non-sandbox computer (this device). Subtasks on cloud computers need none.

{
  "status": "partial",
  "done": 1,
  "subtasks": 2,
  "max_parallel": 2,
  "elapsed_ms": 4310,
  "cost_usd": 0,
  "results": [
    { "subtask": 1, "computer": "cmp_build_box", "status": "done", "steps": ["…"] },
    { "subtask": 2, "computer": "cmp_report_box", "status": "escalated", "reason": "…", "screen": {} }
  ]
}

​

Best-of-N

{
  "member": "run_subtask",
  "input": {
    "goal": "Book the 9:00 meeting room for Friday",
    "success": [{ "text": "Booked" }],
    "best_of": 3,
    "computers": ["cmp_cloud_1", "cmp_cloud_2", "cmp_cloud_3"],
    "max_cost_usd": 0.25
  }
}

1. Rollouts. One rollout per listed computer, all at once (best_of 2–5, at least that many distinct ids in computers). Every computer must be a sandbox/cloud computer. Listing this device, or any computer that isn’t a sandbox, refuses the call with 409 sandbox_required. Rollouts never run as copies on the person’s own machine.
2. Narratives. Each rollout’s step trace becomes a narrative: every step’s chosen option and what happened after it (ran, the screen changed, ran, the screen didn't change, failed: …), whether the success checks held, and what the final screen shows.
3. Shortlist. The candidates are the rollouts whose success checks held. If none held, the candidates are the rollouts that ended done. If exactly one rollout is a candidate, it wins with no judge call.
4. Judge. With several candidates, one decision of kind judge over their narratives (POST /v1/decisions, options r1…rN) picks one. The fast tiers answer first. When they abstain, the oracle (the planner-class tier) gets one try, but only while the call is under max_cost_usd. If nobody picks, the call returns escalated with every narrative so the planner decides.

{
  "status": "done",
  "best_of": 3,
  "chosen": { "rollout": "r2", "computer": "cmp_cloud_2", "why": "the judge (local, confidence 0.91) picked r2 by the step narratives among 2 rollouts that reached the goal" },
  "judge_decisions": 1,
  "elapsed_ms": 9120,
  "cost_usd": 0,
  "rollouts": [
    { "rollout": "r1", "computer": "cmp_cloud_1", "status": "done", "narrative": "Rollout r1 on computer cmp_cloud_1: ended done.\n1. click the button \"Book\" -> ran, the screen changed\n…", "result": {} }
  ]
}

​

Budgets

​

API over GUI

​

Related pages

• How computer use decides
• Computer toolset
• Allternit Driver
• Decisions API