Overview
/silver:multi-ai-task is the canonical opt-in primitive for dispatching bounded work to multiple model backends. The host remains the single workflow driver; workers receive neutral work-item manifests and return auditable results for deterministic reconciliation.
Use it when independent perspectives materially improve the result. Ordinary tasks should stay on the smallest safe single-host workflow.
Catalog Composition
WF-SILVER-MULTI-AI-TASK
→ AF-MULTI-AI-TASK
The flow uses the multi-ai-task-v2 spine. It resolves allowed pools for the active host, enforces the caller's budget and concurrency limits, and records dispatch and result indexes.
Execution Contract
- Confirm explicit multi-AI opt-in and a bounded, decomposable task.
- Resolve the active backend pool and apply budget, concurrency, and timeout policy.
- Write neutral work-item manifests and dispatch independent workers.
- Collect results, validate result envelopes, and reconcile deterministically.
- Record evidence before the workflow can claim completion.
Outputs
- Dispatch ledger and work-item manifests
- Per-worker result envelopes and result index
- Budget and execution evidence
- Deterministically reconciled final result