Amazon has upgraded the Bedrock AgentCore runtime for faster and cheaper AI agent operation
Amazon has announced an update to the Bedrock AgentCore runtime: according to the company, session startup will speed up by up to 15×, capacity will increase by 4×, and operational efficiency will improve by up to 68%, while maintaining isolation in Firecracker MicroVMs.
Amazon has announced an update to the runtime layer of Amazon Bedrock AgentCore, which serves as a fully managed environment for deploying and operating AI agents. According to the company, the new version speeds up agent session startup by up to 15×, increases parallel session capacity by 4×, and improves operational efficiency by up to 68%. Each session continues to run in a hardware-isolated Firecracker MicroVM with its own identity, security policy, and audit trail. Sessions can now use up to 8 vCPUs and 32 GB of memory, support container images up to 10 GB, and startup takes, according to the company, less than 2 seconds regardless of image size or the number of concurrent sessions.
The company is also adjusting its pricing approach to reflect the agent's actual work – the runtime is to release memory within minutes after it is freed by the agent, since, according to the company, agents spend a large portion of their time waiting for a response from a model, a tool, or a human. The rest of the description of the pricing change was not available in the source.
Several companies share their own experiences with migrating to the new version in the text: Coinbase, according to its representative, shortened the agent session cold start from nearly a minute to less than 5 seconds even under high load, and doubled its weekly agent volume. Mission Cloud reports a 35% speedup in cold starts for inference requests, a 77% speedup for health checks, and a 17% reduction in invocation costs. Zuora reports an approximately 60% reduction in memory usage per invocation under increased traffic with stable latency. Rescale states that thanks to the service, it can deploy and scale agents without building its own infrastructure.
Amazon bases this context on data from METR (the duration of tasks that agents can handle independently roughly doubles every four months), McKinsey (40% of large enterprises are now scaling AI agents, up from 27% last year), and Deloitte (only 15% of organizations have scaled multi-agent systems, while 70% say they do not yet trust their agents and do not know how to manage them).
Why it matters
For companies running or planning to deploy AI agents at larger scale, the key factor is the balance between security, response speed, and infrastructure costs – according to the cited customer figures (Coinbase, Mission Cloud, Zuora), the update genuinely reduces cold starts as well as the memory/cost burden per invocation. For developers of agentic systems, this means less work with custom sandboxes and warm pools, since session startup and scaling are now handled directly by the managed service.
Two audiences, two different impacts
What this means
For individuals
Developers and architects building AI agents on AWS can simplify their own scaling solutions and reduce the need for manually tuned sandboxes thanks to faster startups and lower memory requirements.
For a business
Companies running agentic workloads on AWS can, according to documented customer figures (Coinbase, Mission Cloud, Zuora), reduce cold starts, memory load, and invocation costs as agentic traffic volume grows.
DevelopmentCheck the original
Event sources
clearly official source · 1 publisher, 0 independent. We count feeds from the same owner only once.