AWS publishes a guide for a multi-agent engineering-decision assistant using Claude models on Bedrock AgentCore
Amazon Web Services described the architecture of a multi-agent system using Claude models and Amazon Nova Multimodal Embeddings for evaluating aerospace component substitutions. The decision is always made by an engineer, with the system only preparing supporting materials.
Amazon Web Services published on its blog a guide for building a multi-agent assistant that helps engineers evaluate substitutions for qualified aerospace components – a task that, according to the article, typically takes days of going through supplier specifications, interface control documents (ICDs), and qualification records. The architecture uses three agents based on Claude models: a planner (Claude Haiku 5.5) decides how to split the query, specialists (Claude Sonnet 5.5) search a multimodal knowledge base and evaluate candidates according to deterministic tolerance rules, and a supervisor (Claude Opus 5.5) then synthesizes the results from multiple specialists into a response.
The system is built on the Amazon Bedrock AgentCore platform, the open-source Strands Agents SDK, and the Amazon Nova Multimodal Embeddings model, which enables search across text, photographs of components, and charts from datasheets within a single vector space. The knowledge base contains both structured technical data and documents that engineers typically read – ICDs, photographs, and scanned sheets with performance curves.
A key design principle, according to the company, is that the final decision is always made by a human – the system only gathers evidence and prepares an Integration Impact Assessment document, and never approves the substitution itself. It therefore combines language models with deterministic code wherever reliable, repeatable rule-based scoring is needed.
The complete code is available as a set of Jupyter notebooks on GitHub. The remaining technical details of the architecture and setup were not available in the processed excerpt of the article – see the source article for details.
Why it matters
For companies working with qualified or safety-critical components (aerospace, defense, manufacturing), the guide demonstrates a possible way to shorten manual review of specifications when searching for replacement parts while maintaining a formal, auditable decision trail – without shifting final approval to the AI. For developers and architects of AI systems, it is a concrete example of how to combine multiple tiers of Claude models (routing, retrieval, synthesis) with multimodal embeddings and deterministic scoring within a single agentic architecture on Bedrock AgentCore.
Two audiences, two different impacts
What this means
For individuals
Developers and engineers building AI agents gain a concrete reference architecture for how to divide tasks across different tiers of Claude models (Haiku 5.5, Sonnet 5.5, Opus 5.5) and combine them with multimodal retrieval and deterministic code.
For a business
Companies with complex catalogs of qualified components can use this architecture as a default template to reduce manual work when evaluating part substitutions, with final approval still resting with a human – however, this is a published reference guide, not an announced deployment at a specific customer.
ProcessesCheck the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.