Feature request: batch-spawn forked subagents from a shared preloaded context #18683

Please add an API / runtime primitive for batch-spawning multiple forked subagents from one shared preloaded context.

Today, fork_context=true appears to fork from the current caller agent's context, which is useful, but it is still fundamentally a one-agent-at-a-time interface.

What seems to be missing is a higher-level primitive like:

• preload one shared prompt/context once
• fork N subagents immediately after that shared prefix
• provide each fork with a small per-agent payload, such as an agent id, worker index, or task selector
• let each subagent follow the same common instructions but branch by its own identity or assigned shard

Conceptually, this is closer to a SIMD / CUDA-thread style launch:

• same code / same shared prompt prefix
• different thread / worker ids
• different target objects or shards

Why this would help

This would likely improve orchestration efficiency for multi-agent workloads:

• better shared-prefix reuse
• higher KV-cache hit rate
• potentially better MoE expert reuse / routing locality
• lower repeated prompt overhead
• lower latency for fan-out workloads
• cleaner programming model for planner-worker or tree-search patterns

Right now, if I want a whole group of subagents to share the same long setup prompt, I effectively need to recreate that launch pattern one agent at a time. That feels less efficient than a native batch-fork primitive.

Proposed interface shape

Not prescribing exact naming, but something in this family:

spawn_agents({
  shared_context: ...,
  forks: [
    { id: "worker-0", message: "task selector: 0" },
    { id: "worker-1", message: "task selector: 1" },
    { id: "worker-2", message: "task selector: 2" }
  ]
})