OpenRouter Launches Subagent Tool for Task Delegation

ME AI message, according to monitoring by Beating, OpenRouter has launched a server-side agent tool, `openrouter:subagent`, and opened it for testing. It enables large models to delegate independent subtasks during content generation to smaller, cheaper, and faster candidate models. When the main model encounters self-contained tasks that do not require its full capabilities—such as document summarization, structured data extraction, template drafting, and text formatting—it can invoke the agent tool by providing a task name (task_name) and task description (task_description). The delegated subtasks are executed by the working model and returned to the main model as outcomes for subsequent integration. The working model can be any model supported by OpenRouter, specified either by the `parameters.model` field in the tool definition or, if not set, inherited directly from the main model. To enhance execution capability, the working model can also be equipped with independent OpenRouter server-side tools—such as web search (`openrouter:web_search`) or web scraping (`openrouter:web_fetch`)—to perform multi-step reasoning and data retrieval within a sandbox environment before generating the final text. Since working models execute on the server side, they do not support custom function tools requiring client-side executors. As working models cannot access the main model’s conversational context or share memory across tasks, the main model must provide complete background information and output format requirements within the task description. To prevent infinite recursion and uncontrolled costs from nested model calls, OpenRouter has implemented dual safeguards: self-referencing is prohibited in the tool definition, and nesting depth is strictly limited via request headers, with agent tools forcibly stripped from subtask calls. Additionally, a hard cap is enforced on the total number of task executions within a single API request. (Source: BlockBeats)