Strands AWS Advisory Pipeline

• 

  Strands AWS Advisory Pipeline - 4-Agent Architecture Diagram

Inspiration## Inspiration

Prompting LLMs directly often yields inconsistent outputs, unhandled risk factors, or missed context. We built the Strands AWS Advisory Pipeline to create a structured, zero-friction agent governance architecture that cleans, optimizes, executes, and audits user requests before final response delivery.

What it doesThe platform runs a sequential 4-agent workflow leveraging Strands SDK and AWS Bedrock:

• Analyzer Agent: Audits initial prompts for intent, context gaps, and sensitive terms using Pydantic typing.
• Optimizer Agent: Dynamically injects system instructions, runtime constraints, and contextual scaffolding.
• Executor Agent: Dispatches high-precision payloads to AWS Bedrock (anthropic.claude-3-sonnet-20240229-v1:0 in us-east-1).
• Quality Control Agent: Audits output text for hallucination risk, compliance, and assigns a final APPROVE or REJECT verdict.

How we built it*Framework:* Python, strands-agents, pydantic, boto3.

• LLM Engine: AWS Bedrock (anthropic.claude-3-sonnet).
• Resilience Layer: Built-in zero-credential simulation fallback engine so the multi-agent pipeline operates seamlessly without breaking CLI/API workflows during demonstrations.

Challenges we ran intoEnsuring smooth state inheritance between 4 sequential agents without introducing schema errors. Injecting automated security overrides when flagged terms like password are detected while keeping system latency under 600ms was also a core focus.

Accomplishments that we're proud of - Successfully built a fully functional 4-Agent pipeline (Analyzer, Optimizer, Executor, Quality Control) that coordinates seamlessly using Strands SDK and Pydantic schemas.

• Engineered a robust Zero-Credential Bedrock Simulation Engine that fallback-executes during testing without requiring active AWS API keys or breaking the CLI interface.
• Built-in real-time safety auditing that flags critical risk patterns (like sensitive credentials/passwords) and forces safety compliance.

What we learned We learned how to combine deterministic Pydantic schemas with agent execution logic to build trustworthy, production-grade LLM applications on AWS.

What's next for Strands AWS Advisory Pipeline - Adding parallel multi-agent evaluation.

• Integrating native AWS Guardrails directly into the Quality Control agent.

Built With

• ai-agents
• aws-bedrock
• claude-3
• pydantic
• python