Agents that spike once an hour and idle the rest of the day were paying for their peak around the clock. AWS says the new AgentCore runtime fixes that.
Released Friday, the reworked managed compute layer under Bedrock AgentCore reclaims memory as sessions finish with it, instead of holding an allocation from start to finish. The old behavior meant long-running or bursty agents carried their high-water mark well past the moment demand fell.
Cold starts got the second pass. A session that lands on an already initialized environment is running in under 100 milliseconds, but warm environments cost reserved compute. Most sessions instead boot fresh, pulling the container image and initializing the agent, and that wait scales with image size and concurrency.
AWS isolated the platform’s own contribution by testing an empty echo agent that calls neither a model nor a tool. A Python client in us-west-2 called agents in us-east-1 over the public internet through the boto3 SDK, with 5,000 cold invocations per agent across both runtime versions and five image sizes.
Three findings came out of that run. The new runtime held a P75 cold start near 2 seconds whether the image was 200MB or 2GB. The original climbed from about 5.4 seconds to nearly 30 across the same range. And the agent’s own code needed only about 34 milliseconds at P75, so almost all of the measured time belonged to platform startup.
Pricing still tracks usage down to zero, with no charge for idle CPU waiting on input. AWS suggests opening a session the moment a user engages, so the environment warms while they type.
