Orchestrator mode: plan → todo → delegate each step to a subagent → integrate summaries → finish.
Prompt-driven delegation inside deterministic runtime rails — the 2026 consensus across Claude Code, Anthropic's multi-agent research system, OpenAI's manager pattern, and LangGraph's supervisor:
- Depth 1 — workers never spawn workers (enforced in SpawnAgent).
- Fan-out ≤ provider slots —
max_concurrencyis capped at the provider's request-queue depth, so parallel workers can never exceed what the GPU actually serves. - Strict task briefs —
assigns[:strict_spawn]makes SpawnAgent reject delegation without the four-field brief (objective, expected_output, tool_guidance, boundaries). - Summary-only returns — workers' final messages are the only thing entering this context (and the prompt forbids restating them).
Mechanically it follows the PlanAndSolve shape: one tool-free planning
iteration, then delegation-driven ReAct. The worker contract rides in as
assigns[:subagent_prompt_suffix] (appended by SpawnAgent to every
worker's system prompt) unless the caller supplied their own.