Introduces the Agent-as-a-Tool paradigm, a practical architecture that addresses Tool Space Interference (TSI) in LLM agents by using a RAG-based Agent Bank to dynamically discover and assemble stateful sub-agents at runtime. Built on Google ADK and TypeScript, the system implements semantic search to retrieve only the minimum required agents per task, supports parallel and sequential execution strategies, enforces Zero-Trust security via ephemeral in-memory runners, and includes Human-in-the-Loop checkpoints for destructive operations. A complete Node.js/TypeScript implementation with three sample agents (currency, weather, Google Workspace) is provided, demonstrating both web server and A2A server deployment modes.
Table of contents
AbstractIntroductionRepositoryThe Agent-as-a-Tool ParadigmArchitecture and Execution WorkflowSystem Execution Flow in PracticeProject Setup and PrerequisitesSecurity Considerations and Zero-Trust GovernanceFuture Perspectives on Dynamic OrchestrationSummaryExplanatory videoSort: