- Subject
- Windows & PowerShell
- Filed
- Source
- James Smith
- Loans
- 379
AWS AgentCore Runtime Adds 8-vCPU Sessions, Scale-to-Zero and Baseline Pricing
AgentCore sessions can now use up to 8 vCPUs and 32GB of memory. Teams should weigh scale-to-zero savings against resume latency and baseline costs.

AWS has expanded the compute and memory options in Amazon Bedrock AgentCore Runtime, giving developers more room to run demanding agent tasks inside isolated sessions. Each session can now use up to 8 virtual CPUs (vCPUs) and 32GB of memory, up to four times the compute capacity of the previous version. Container images can reach 10GB, and startup times are under two seconds regardless of image size, according to AWS.
For enterprise teams, the change could reduce the need to split workloads across sessions or maintain idle capacity to avoid startup delays. But lower runtime resource use does not remove every cost or operational concern: storage, model calls and external tools may still incur charges, while agents that scale to zero need careful resumption checks.
More room for complex agent workloads
The larger session limits are aimed at workloads that can outgrow smaller execution environments. Examples include compiling large codebases, browser agents that load full web pages, and running a small model locally. With up to 8 vCPUs, 32GB of memory and 10GB container images, these tasks can fit within AgentCore’s isolated runtime, according to Ashish Chaturvedi, executive research leader at HFS Research.
That capacity may change how developers divide work. Chaturvedi said developers had needed to break up some tasks or run them outside the isolated environment when earlier session limits proved insufficient. The expanded limits could keep more of that work within the runtime boundary.
Startup speed is another part of the update. Chaturvedi said that when a new isolated session takes nearly a minute to start, teams may keep pre-provisioned capacity ready to avoid delays, even when agents aren’t using it. With startup times under two seconds, he said, developers can create isolated sessions on demand rather than holding warm capacity in reserve.
Idle sessions can release memory
AgentCore Runtime can reclaim unused memory within minutes after an agent releases it, and idle agents can scale to zero. AWS says this can reduce memory consumption per session by up to 68% for multi-turn agents with natural pauses between interactions.
The feature is particularly relevant to agents that wait for model responses, tool calls or human input. Their resource needs can vary across a session: an agent may need substantial capacity while executing code, then sit idle while waiting. Stephanie Walter, practice leader of AI stack at HyperFRAME Research, said reclaiming memory and scaling idle sessions to zero could lower the cost of keeping long-running agents available.
Chaturvedi said the runtime’s behavior could also spare developers from writing their own logic to shut down inactive sessions, save state and restart agents. That may simplify long-running, multi-turn workflows, although teams still need to confirm that agents resume correctly.
Baseline pricing targets always-on workloads
AWS has added baseline pricing for always-on agents. The model sets a minimum memory allocation for each session and lets workloads burst to as much as four times that baseline when they need extra capacity. AWS says this gives enterprises a way to keep sessions available without paying for peak capacity upfront, while retaining room to handle higher demand.
The cost outcome depends on how closely the baseline matches actual usage. Chaturvedi cautioned that a baseline set too high can leave enterprises paying for unused capacity across sessions. A baseline set too low may lead to frequent bursts and less predictable costs. He said the model is likely to suit workloads with relatively stable resource requirements; usage-based pricing may fit unpredictable or experimental workloads better.
Teams considering the pricing option will need to understand how burst capacity is billed and compare the baseline with observed workload patterns. Monitoring after deployment can show whether the chosen allocation fits real demand.
Check resumption, access and total cost
Scaling an agent to zero can cut runtime resource consumption, but it doesn’t automatically eliminate costs for storage, model calls, external tools or other services. Walter said developers should also test what happens when an agent resumes: whether its files remain available, its credentials are still valid, and it can determine whether its previous action completed.
That last check matters for workflows that can make consequential changes. If a resumed agent can’t tell whether it finished its prior task, it could repeat an action such as a payment or a production change. Teams should validate state handling and resumption behavior against their own workflows before relying on scale-to-zero for them.
There’s also an access management consideration. Chaturvedi warned that cheaper idle sessions could encourage teams to leave agents running indefinitely, even when they’re no longer needed. Those agents may retain permissions and access to enterprise resources, so regular reviews of deployments and agent access remain necessary.
Scale-to-zero may also add latency when an agent needs to resume, making it a poor fit for some latency-sensitive tasks. The practical choice depends on how much idle time a workload has, how quickly it must respond, and how reliably it can restore its state.
Availability across 19 AWS Regions
AWS says the updated AgentCore Runtime capabilities are available across 19 Regions: N. Virginia, Ohio, Oregon, Canada, São Paulo, Frankfurt, Ireland, London, Milan, Paris, Spain, Stockholm, Malaysia, Mumbai, Seoul, Singapore, Sydney, Thailand and Tokyo.
For IT teams, the next step is to match the new session limits, memory behavior and pricing model to specific agent workloads. Workloads with long pauses may benefit from memory reclamation, while always-on tasks may warrant a baseline allocation. Testing startup, resumption, access and burst charges will show whether those options fit each deployment.