According to Beating, OpenRouter has launched the server-side agentic tool 'openrouter:subagent' in beta testing, enabling large language models to delegate subtasks to smaller, cheaper, and faster alternative models during content generation. When the primary model encounters self-contained tasks that don't require its full capabilities—such as document summarization, structured data extraction, template drafting, and text formatting—it can invoke the tool by specifying the task name and description. The delegated task is executed by a worker model, which returns the result back to the primary model for integration.
Worker models can be any model supported by OpenRouter and inherit the primary model if not explicitly specified. They can access server-side tools like web search and web scraping for multi-step reasoning before generating final output. To prevent infinite recursion and cost overruns from nested calls, OpenRouter has implemented dual safeguards: prohibiting self-references in tool definitions and enforcing maximum nesting depth via request headers, while also capping the total number of task executions per API request.
Related News