Someone finally locked down recursive calls—double-layered protective nesting is essential; I've seen too many horror stories of AI agents looping infinitely before.

OpenRouter launches server-side proxy tool openrouter:subagent, with task decomposition during testing: the main model dispatches sub-tasks to working models for execution. The main model schedules based on task name and description, sub-tasks are completed by working models, which can choose any model supported by OpenRouter; if not set, they inherit from the main model. Working models can be configured with independent tools and perform multi-step reasoning and data retrieval before text generation. Working models do not access the main model's context and must provide complete background in the task description. To prevent infinite recursion, the system has dual protections, prohibiting self-reference, limiting nesting depth, and imposing a hard cap on total executions.