Bedrock agent orchestration with tool-use and guardrails

The problem

A team had a working chat prototype but no way to take real actions in their systems. Prompts were hand-tuned per use case, tools were defined inline, and no one trusted the model to write or send anything. Bedrock was chosen for data residency, but they needed a real orchestration layer on top.

The solution

Built an orchestration layer on Bedrock. A planner agent picks a route, specialist agents own bounded skills, and tools are typed JSON-schema contracts validated on both sides. Bedrock Guardrails block sensitive content and PII. Sensitive actions enter a human-approval queue with full reasoning trace. Every run is logged with prompt, tool calls, outputs, and cost, queryable for eval and replay.

Architecture

Stack

Outcomes