shared-subagent-batching

Sub-agent batching orchestrates parallel per-unit sub-agents by launching up to a configured batch size simultaneously. Each sub-agent processes exactly one unit, ensuring deterministic prompts and outputs before draining the worklist.

Sub-agent retries work by validating each unit's JSON output upon completion and retrying failed units up to three times. After draining the worklist, a final aggregate validation runs to ensure all units meet consumer-specific requirements.

The best way to coordinate multiple sub-agents with deterministic outputs is using a launch loop with a fixed batch size. Source sub-agent prompts from a template file, process exactly one unit per call, and validate outputs after draining the worklist.

Yes, you can use a custom prompt template for each sub-agent batch. Sub-agent prompts are sourced from each consumer skill's references/subagent-prompt-template.md file, ensuring deterministic instructions for every unit processed in the parallel batch.

Yes, each parallel sub-agent inherits the main agent's model automatically. This ensures consistent behavior across all concurrent sub-agents running within a batch, while processing exactly one unit per sub-agent call for deterministic outputs.

You should not use batching for sub-agent orchestration when processing a single unit or when strict sequential ordering is required. Batching is designed for multi-unit workflows needing parallel concurrency with validation and retries after draining the worklist.