Descriptions:
Elizabeth Fuentes Leon, AWS developer advocate, presents five code-level techniques to cut AI agent hallucinations at the AI Engineer conference. The central argument is that prompt tweaks alone are insufficient — each technique requires a direct architectural change to how the agent operates.
The five methods are: semantic tool selection (embedding tools in a vector database and filtering to the three most relevant per query, dropping token usage from ~3,000 to under 300); graph RAG (replacing vector retrieval with structured Cypher queries via Neo4j for aggregation and multi-hop reasoning tasks); multi-agent validation (a second agent reviews every response before it reaches the user); neuro-symbolic guardians (hard Python rules the model cannot override regardless of prompt instructions); and runtime guardians (self-correcting steering that redirects the agent without issuing a hard stop or surfacing an error to the user).
All demos are built around a travel agent powered by the Strands framework — AWS’s open-source agentic toolkit — and run in a Jupyter notebook using OpenAI models alongside local sentence-transformer embeddings for the tool-vector database. Each technique is shown side-by-side with and without the change, giving developers a clear before-and-after view of failure modes and their fixes. Code and setup instructions are available via QR code in the talk.
📺 Source: AI Engineer · Published July 11, 2026
🏷️ Format: Hands On Build
