Simulating an Expert Team with Agentic AI and Amazon Bedrock AgentCore

Complex technical questions often take more time to assemble and coordinate the right experts than to actually find the answer. To address this, AWS built a coordinator-based dynamic swarm pattern on top of Amazon Bedrock AgentCore, combining 7 domain expert agents with a coordinator.

The coordinator first checks past conversations in AgentCore Memory to handle follow-up questions, and only convenes the expert team when necessary. Agent code runs on AgentCore Runtime, tool access happens through AgentCore Gateway and MCP (Model Context Protocol), and actual query processing is handled by AWS Lambda and Amazon Bedrock Knowledge Bases.

Knowledge is separated by domain. The HPC, Quantum, Visual Computing, Spatial Computing, IoT, Technology Partnerships, and Emerging Technologies agents each look up documents in their own area, generate answers using an agentic RAG approach, and then hand off to each other to enrich the content. The system imposes limits of up to 20 handoffs, 20 iterations, and a total of 30 minutes per swarm.

There are three example queries. A single-domain question, like explaining AWS PCS, is handled by one expert. A digital twin question involving 2,800 cameras and sensors deployed across a 46,000-square-meter factory is answered jointly by the IoT, Spatial Computing, and Generative AI agents. And an HPC weather simulation question demonstrates network-style collaboration, where agents validate each other's points and iteratively refine their proposals.

The core value lies not in producing a finished final answer, but in generating a draft within minutes that a human expert can then verify and refine. Compared to the broad, safe generalities offered by a typical direct LLM call, this system provides a more concrete starting point that includes combinations of AWS services and implementation patterns.