<p>AWS described a multi-agent engineering assistant for aerospace part-substitution review, built on Amazon Bedrock AgentCore runtime, Amazon Nova Multimodal Embeddings and the open-source Strands Agents SDK. The stated problem: when a qualified component goes obsolete or hits a long lead time, an engineer spends days cross-referencing vendor datasheets, interface control documents and qualification records before documenting a justification, and an incorrect substitution on a flight-critical part can cause mission failure.</p><p>The architecture is a planner that routes work across three specialists, with a supervisor that synthesises results when more than one specialist runs. The planner uses Claude Haiku 5.5, specialists use Claude Sonnet 5.5 and the supervisor uses Claude Opus 5.5. One specialist searches a multimodal knowledge base that indexes diagrams, thermal plots and datasheet charts that text-only retrieval misses; another scores candidate parts against deterministic tolerance rules; a third grounds lifecycle claims in live web data. Part photographs are compared as part of the evidence assembly.</p><p>The design rule stated in the post is that the engineer makes every decision and the system assembles the evidence for it. That rule decides where large language models are used and where deterministic code is used instead, and the system drafts documents without ever approving a substitution on its own.</p>
Our takePutting tolerance checks in deterministic code and leaving retrieval and drafting to the models makes the pass/fail path auditable, which is the right trade for flight-critical parts. The new failure surface moves to the boundary: a misparsed datasheet value or a unit mismatch inside the rule still yields a confident wrong score, and the agent presents it as evidence. Lifecycle claims from live web data need source dating to be checkable.