AWS shows how to cut idle costs when AI agents process requests

AWS's guidance addresses a fundamental cost asymmetry in serverless AI agent workflows. Amazon Bedrock AgentCore uses a consumption-based model that charges only for memory (and model output tokens) while the agent is idle—it does not bill CPU time during waits on language model generation or external tool calls. The compute service that called the agent, however, has no such behavior: a Lambda, container, or EC2 instance that blocks on a synchronous call holds its entire allocation and incurs billing for every second of that idle time. This mismatch means the caller's cost tracks the agent's runtime, not the agent's actual work. The three patterns AWS describes—task-token callback, direct service integration, and durable functions—all follow the same principle: start the agent, release the caller's compute immediately, and resume only when the agent has a result. The differences are trade-offs in implementation: task-token callback keeps a Lambda in the path for custom logic; direct service integration removes the Lambda entirely by integrating Step Functions directly with AgentCore; and durable functions allow the same orchestration to be expressed as code in a single Lambda rather than a state machine.