A validated multi-agent orchestrator, built on AWS

A SaaS B2B business working with Cloud Combinator on AWS. The client is anonymised at their request.

The proof of concept had four focus areas, each essential to making agentic AI dependable enough for a product.

A request arrives through Amazon API Gateway and invokes the Orchestrator, an AWS Lambda function. It builds a compact execution plan, consults the Agent Registry (Amazon DynamoDB with a Lambda accessor) to pick the best agent for each step, and validates every inter-service payload against pydantic models and JSON Schemas. When needed, an Evaluator Lambda applies its agent and model ladders and returns a verdict with proposed confidence. For visual output, an inline or dedicated plotting agent produces a Plotly or Highcharts specification, and a Rendering Lambda validates it and exports PNG, SVG or CSV to Amazon S3, returning presigned URLs. Every response carries a concise narrative, validated tables or charts, and two suggested next actions.

Acceptance was tied directly to the success criteria: two representative flows executed end to end, including memory retrieval and plotting or export, with observable agent selection, validated contracts and accessible artifact links. The proof of concept delivered a serverless, least-privilege agentic platform that demonstrated AWS is a suitable foundation for taking the capability forward.

• End-to-end orchestration — A request flows from API Gateway through the Orchestrator, which executes a plan, validates every exchange, and returns a narrative, tables or charts, and two next actions.
• Registry-driven selection — The Orchestrator selects a single best agent per step from the DynamoDB-backed registry, with choices that are observable and reproducible in traces.
• Evaluator with ladders — The Evaluator applies agent and model ladders and returns a verdict and proposed confidence, which the Orchestrator combines with deterministic checks into a final confidence.
• Grounded memory and guardrails — AgentCore and Bedrock Knowledge Bases on OpenSearch Serverless supply tenant-aware retrieval, with Bedrock Guardrails enforcing policy on ingress.
• Observability by default — CloudWatch dashboards and alarms and AWS X-Ray traces span the Orchestrator, Registry, Evaluator, agents and rendering, tracking selection distribution, confidence deltas, Knowledge Base hit-rate and validation pass-rate.
• Handover — Versioned contracts, CloudFormation, documentation and recorded walkthroughs were delivered, with two representative flows signed off end to end.

With a validated, observable agentic foundation proven on AWS, the client has a clear path to integrate the orchestrator into their platform and harden it towards production, extending the agent registry, tuning the evaluator ladders, and scaling the knowledge and memory layers as usage grows.