AI agent hallucinations occur when agents fabricate statistics, pick wrong tools, ignore business rules, or claim success when operations fail. This workshop provides 5 hands-on techniques — Graph-RAG, semantic tool selection, multi-agent validation, neurosymbolic guardrails, and agent steering — plus a production deployment demo on Amazon Bedrock AgentCore.
Based on the Dev.to series Stop AI Agent Hallucinations: 4 Essential Techniques and 5 Techniques to Stop AI Agent Hallucinations in Production.
Built with Strands Agents and Amazon Bedrock. The same patterns apply in LangGraph, AutoGen, CrewAI, or any other agent framework.
| Approach | Hallucination Risk | Retrieval Method | Best For |
|---|---|---|---|
| Standard RAG (vector) | High — returns similar content even when irrelevant | Cosine similarity | General Q&A |
| Graph-RAG (Neo4j) | 73% lower — grounded in entity relationships | Graph traversal + Cypher | Structured domains (hotels, products, finance) |
Key insight: Vector search always returns something similar, even when the answer doesn't exist in the database — causing fabrication. Graph-RAG returns only what's explicitly connected in the knowledge graph.
| # | Demo | What It Solves | Key Result | Stack |
|---|---|---|---|---|
| 01 | Graph-RAG vs RAG | Fabricated statistics, incomplete retrieval, out-of-domain hallucination | 73% fewer hallucinations with knowledge graphs |   |
| 02 | Semantic Tool Selection | Wrong tool picks, token waste at scale (29 tools) | 89% token reduction, higher accuracy |  |
| 03 | Multi-Agent Validation | Undetected hallucinations, fabricated responses | Executor-Validator-Critic cross-check pipeline |  |
| 04 | Neurosymbolic Guardrails | Agents ignoring business rules in prompts | Symbolic rules enforced via lifecycle hooks |  |
| 05 | Agent Control Steering | Hard-blocking stops the task instead of fixing it | Agent self-corrects instead of failing |  |
| 06 | Amazon Bedrock AgentCore | Taking all 5 techniques to production on AWS | Hotel booking agent on Amazon Bedrock AgentCore with Lambda + DynamoDB |    |
Each demo builds on the previous one. You can run any demo independently, but the learning path is designed to be progressive:
Phase 1 — Understand Failures: Demos 01-02 show why agents fail — RAG hallucinations and tool selection errors backed by academic research.
Phase 2 — Detect & Prevent: Demos 03-04 show how to catch failures — multi-agent validation and symbolic rule enforcement.
Phase 3 — Self-Correct: Demo 05 shows how to steer agents to self-correct instead of blocking (Agent Control).
Phase 4 — Deploy to Production: Demo 06 shows how to ship all 5 techniques to production on AWS using Amazon Bedrock AgentCore, DynamoDB, and Lambda.
- Python 3.9+
- uv package manager
- AWS account with Amazon Bedrock access (Claude Sonnet 4 enabled in your region)
cd 01-graphrag-demo # or any demo folder
uv venv && uv pip install -r requirements.txt
# Run the demo
uv run <main_script>.py
# Or open the .ipynb notebook in your IDE (VS Code, Kiro, or any editor with notebook support)Each demo README has specific setup instructions and prerequisites.
Demo 01 requires a Neo4j instance to build and query the knowledge graph.
Running as part of a workshop: A Neo4j Aura instance will be provided. You will receive the connection credentials (NEO4J_URI, NEO4J_USERNAME, NEO4J_PASSWORD) at the start of the session — add them to a .env file in 01-graphrag-demo/.
Running independently: Create your own free Neo4j Aura instance:
- Go to console.neo4j.io and create a free AuraDB instance
- Download the credentials file when prompted — it contains your URI, username, and password
- Create a
.envfile in01-graphrag-demo/:NEO4J_URI=neo4j+s://<your-instance>.databases.neo4j.io NEO4J_USERNAME=neo4j NEO4J_PASSWORD=<your-password> - Enable the APOC plugin in your Aura instance settings
- Run
build_graph_lite.py(30 docs, ~15 min) orbuild_graph.py(300 docs, ~2 hours) to populate the graph
This repository addresses four main categories: (1) fabricated statistics — when RAG agents guess numbers instead of computing them, (2) wrong tool selection — when agents pick inappropriate tools from large toolsets, (3) business rule violations — when agents ignore constraints expressed only in prompts, and (4) undetected failures — when single agents claim success without validation.
Yes. The patterns (Graph-RAG, semantic tool filtering, multi-agent validation, neurosymbolic guardrails, steering controls) are framework-agnostic concepts. These demos use Strands Agents, but the same approaches apply in LangGraph, AutoGen, CrewAI, Haystack, or custom implementations. The key insight is architectural, not framework-specific.
Yes. All demos use Amazon Bedrock (Claude Sonnet 4) as the default LLM provider. You need an AWS account with Bedrock access enabled in your region.
Demos 02-05 run in under 5 minutes. Demo 01 has a lite mode (30 docs, ~15 minutes) and full mode (300 docs, ~2 hours) for building the knowledge graph.
All demos default to Amazon Bedrock (Claude Sonnet 4) but work with any provider supported by Strands Agents: Anthropic API, OpenAI, Ollama (local models), or any OpenAI-compatible endpoint. See Strands Model Providers for configuration.
Bedrock access denied: Ensure the model (us.anthropic.claude-sonnet-4-5 or similar) is enabled in your region via the Bedrock Model Access console.
Neo4j connection fails (demo 01): Verify NEO4J_URI, NEO4J_USERNAME, and NEO4J_PASSWORD are set in your .env file and that APOC is enabled on your Aura instance.
OpenTelemetry warnings: "Failed to detach context" warnings in demos 03-05 are harmless and do not affect functionality.
Model alternatives: Change the model in any demo by modifying the BedrockModel(model_id=...) call. See Strands Model Providers for all supported options.
For demo-specific issues, check the troubleshooting section in each demo's README.
- Fuentes, E. (2025). Stop AI Agent Hallucinations: 4 Essential Techniques. Dev.to / AWS. dev.to/aws/stop-ai-agent-hallucinations-4-essential-techniques-2i94
- Fuentes, E. (2025). 5 Techniques to Stop AI Agent Hallucinations in Production. Dev.to / AWS. dev.to/aws/5-techniques-to-stop-ai-agent-hallucinations-in-production-oik
- Edge et al. (2024). From Local to Global: A Graph RAG Approach to Query-Focused Summarization. Microsoft Research. arXiv:2404.16130
- AWS. Reducing hallucinations in large language models with custom intervention using Amazon Bedrock Agents. AWS Blog
Contributions are welcome! See CONTRIBUTING for more information.
If you discover a potential security issue in this project, notify AWS/Amazon Security via the vulnerability reporting page. Please do not create a public GitHub issue.
This library is licensed under the MIT-0 License. See the LICENSE file for details.
Last updated: March 2026 | Strands Agents 1.27+ | Python 3.9+