Inside Claude 3.7: The Agent Dispatch Mechanism

The latest Anthropic blog post references a tool called dispatch_agent, which appears to facilitate multi-step searches. Through reverse-engineering the generated JavaScript, we see it involves launching new agents with prompts that specify tool usage patterns. This includes scenarios such as searching for specific keywords or file paths, using specialized tools to improve speed and precision.

The agent usage notes emphasize concurrency, recommending multiple simultaneous invocations to improve performance. The agent then returns a single message containing its final report, invisible to the user unless explicitly summarized. To ensure autonomy, every detail of the desired search or retrieval must be described in the prompt since no further interaction is possible once the agent starts.

Interestingly, the system configures “tengu_thinking” tokens to manage the agent’s depth of processing. Phrases like “think harder” or “think more” trigger different token budgets, allowing more extended reasoning or solution searches. If those phrases are absent, the system defaults to zero tokens, indicating minimal extra thought.

Notably, the agent can’t directly modify files—tools like ${G5.name} or ${RI.name} are off-limits within the agent. Instead, the agent is restricted to searching, reading files, or summarizing. This design reflects a careful approach to minimize context bloat, enforce boundaries, and optimize performance while still enabling detailed multi-step tasks.

More information: https://www.anthropic.com/research/building-effective-agents