AWS describes an approach for routing coding agent requests across multiple models on Amazon Bedrock. The post argues that the relevant metric for coding agents is cost per solved task rather than price per token, since agents operate in terms of tasks.
The walkthrough uses the AWS CDK to deploy LiteLLM together with the NVIDIA NeMo Switchyard routing plugin. It then measures cost per solved task for a routed model group and compares those results against single-model baselines.
Why it matters
Routing requests to different models can affect the overall cost of completing agent tasks. By framing the comparison around cost per solved task, the post offers a way to evaluate routing against relying on a single model.
Who should care
Developers and teams building coding agents on Amazon Bedrock who want to compare routed model configurations to single-model setups.