AWS gives each AI agent session 8 CPUs and 32GB of memory

On Thursday, AWS upgraded the runtime of Amazon Bedrock AgentCore, its tool that operates AI agents for companies. Each session has a maximum of four times more processor power: eight virtual CPUs (processor units) and 32GB of memory. AWS says that agents that wait can scale to zero. AWS also added baseline pricing for agents that operate all the time.

What is new

The runtime gives back memory that an agent does not use, in minutes. Baseline pricing sets a minimum memory for each session. A session can use four times that quantity. Developers can keep agent work in filesystems, such as Amazon S3 storage.

The numbers

AWS says memory use for each session falls by a maximum of 68% for agents that stop between turns. Analyst Ashish Chaturvedi says a session starts in less than two seconds. The source gives no test results from other parties.

What is not known

Analyst Stephanie Walter says scale to zero does not remove the cost of storage, model calls and tools. Chaturvedi warns of a wait when an agent starts again, and of agents that wait but keep their permissions. Walter says an agent that starts again can do a payment again.

Insight: AWS attacks the cost of agent waiting twice

In October 2026, AWS and UC Berkeley released Infinity. This open-source software stops an agent while it waits for a tool. The AgentCore runtime also lets agents that wait scale to zero. This reads as AWS that gives open software to users and sells the runtime. The host with the low cost for agents that wait can win the agent market. For leaders: compare the cost of agents that wait.