On November 12, AWS announced updates to the next generation of its AgentCore runtime. The headline claims are sessions that start up to 15 times faster, four times as much computing power per session, and up to 68 percent less memory used. I’ve read the announcement, the September post that introduced the new runtime, and the pricing page. I haven’t run anything on it.
What the runtime is
Amazon Bedrock is the AWS service companies use to build products on AI models. AgentCore is the part of it for agents, which are AI programs that carry out multi-step work such as reading files, running code, and using a browser. The runtime is where those agents run. Each session gets its own small virtual machine, built with Firecracker, the open-source software AWS also uses to run Lambda, so one user’s session is separated from another’s at the hardware level.
A note on the word memory. In this announcement it means RAM, the working space a computer uses while a program is running. What an agent remembers about you is a separate AWS product, AgentCore Memory.
What makes it next-generation
The new runtime first became available on September 18 in five AWS regions. This week’s release makes sessions larger and adds regions in North America, South America, Europe, and Asia Pacific. Two changes do most of the work.
The first is how sessions start. The old runtime built each one from scratch, which in AWS’s September tests took from 5.4 to nearly 30 seconds depending on the size of the agent’s software. The new one sets the agent up once, takes a snapshot, and restores it for every new session. That took about 2 seconds at every size, which matches the “up to 15x faster” in this week’s post.
The second is idle memory. Agents spend much of a session waiting on a model, a tool, or a person. The old runtime held on to the most memory a session had ever used until the session ended. The new one takes memory back after about two minutes of disuse, and an idle agent drops to zero. AWS says that means up to 68 percent less memory used per session for agents with pauses between turns.
Sessions can also be bigger now, up to 8 virtual processors and 32GB of memory, and an agent can attach storage such as an Amazon S3 bucket from inside its isolated session.
Why AWS built it
AWS’s stated reason is that agents have changed. A chatbot answered one question and stopped. Agents now run for hours, and companies want to run thousands at a time. The announcement cites a Deloitte survey in which only 15 percent of organizations had scaled systems of multiple agents and 70 percent said they can’t yet trust and govern their agents.
The practical reasons are cost and waiting. An agent that needed a lot of memory for a moment kept paying for that peak until its session ended, and a person who waits 30 seconds for a session to start will notice. The September post describes companies building their own workarounds, keeping spare machines ready and paying for them whether or not they were used. My reading is that AWS wants that work and that spending on its platform. The announcement says paying less for the same work “makes the finance conversation that often gates production rollouts easier to close.”
Reading the numbers
The 68 percent is memory used, and AWS hasn’t published a matching figure for the bill. The September post says that on the new runtime “you pay a higher rate but on far fewer GB-hours,” and that for most agents the total goes down. The only cost number in this week’s post is from a customer, Mission Cloud, which measured 17 percent lower cost per call.
One piece is unfinished. The control that ties each session to its own storage folder is promised “by re:Invent,” the AWS conference that starts November 30. As I read it, until then it is up to each company to keep one session’s files apart from another’s.
The runtime is in the registry.