Amazon Web Services is expanding the infrastructure around AI agents, adding faster runtime technology, longer-running compute, centralized agent discovery and payment capabilities as the company builds systems designed to perform multi-step tasks with less continuous human intervention.
The latest development came on September 18, when AWS announced the next generation of Amazon Bedrock AgentCore Runtime, its serverless compute environment for deploying AI agents.
The new Runtime uses serverless microVMs and introduces elastic memory management that can reclaim memory no longer being actively used during a session. AWS said this allows customers to pay for actual memory usage rather than the peak amount allocated to a session.
The system also keeps the serverless model of no pre-provisioning, scale-to-zero operation, hardware-enforced session isolation and usage-based billing.
AWS said the new Runtime improves startup performance by preparing an agent environment once and creating a snapshot. New instances can then restore that prepared environment instead of repeating the complete startup process.
In AWS testing, the new Runtime produced a 75th-percentile cold-start time of 1.9 to 2.0 seconds for container images ranging from 200 MB to 2 GB. AWS reported 5.4 to 30 seconds for the previous version under its test conditions. The company said the test used a simple echo agent to isolate runtime startup performance.
The new Runtime is available in the US East (N. Virginia), US East (Ohio), US West (Oregon), Europe (Ireland) and Asia Pacific (Tokyo) AWS regions. Developers can use the second-generation runtime by setting platformVersion to V2 when creating or updating a runtime.
The September release follows a series of AgentCore additions that show AWS building infrastructure around agents rather than focusing only on the underlying AI model.
On August 6, AWS made AgentCore Runtime Instances generally available. The feature allows agents to run on AWS-managed Amazon EC2 infrastructure while AWS handles provisioning, patching, scaling and lifecycle management.
Runtime Instances provide access to GPU-accelerated, memory-optimized and compute-optimized EC2 instance families. They support agent sessions of up to 14 days, compared with up to eight hours for the standard serverless microVM runtime.
AWS said the capability is intended for sustained, resource-intensive and specialized-hardware workloads. Customers pay for the management of the provisioned compute in addition to their Amazon EC2 costs.
AWS also expanded the infrastructure’s ability to handle larger numbers of simultaneous agent workloads. In July, the company increased default AgentCore runtime quotas to support up to 5,000 active concurrent sessions in US East (N. Virginia) and US West (Oregon), and up to 2,500 in other supported regions. AWS said supported regions also received defaults of 200 agent interactions per second and 25 new sessions per second.
Another part of the expansion is the AWS Agent Registry, which became generally available on August 31.
The registry gives organizations a private, governed catalog for AI agents, tools, skills, MCP servers and custom resources. It provides semantic and keyword search, approval workflows and AWS CloudTrail audit trails.
Organizations can manage registries as code through AWS CloudFormation, Terraform and the AWS CDK. They can also use AWS Resource Access Manager to share registries across accounts and create organization-wide registries.
AWS said the registry can automatically detect agents running on AgentCore Runtime and AgentCore gateways across an organization and add them to a central registry. The service can also be used from Amazon Quick and Kiro IDE, allowing users to find registered agents and tools without manually configuring each connection.
The company is also developing infrastructure that allows agents to transact with other software.
On August 18, AgentCore Payments became generally available. AWS said the feature allows AI agents to discover, access and pay for paid APIs, MCP servers and content.
The service integrates with Coinbase and Stripe Privy wallets, supports configurable payment limits at the infrastructure layer and provides transaction observability through AgentCore. AWS also added support for the Machine Payment Protocol and additional x402 capabilities.
The payments system had originally entered preview in May. AWS said its infrastructure can handle wallet authentication, payment execution, spending governance and transaction monitoring so agents can complete payments during execution.
That capability extends the role AWS is giving agents beyond producing information. An agent can potentially obtain a service, access a paid resource and complete the associated transaction as part of a larger workflow, subject to the controls configured by the developer.
AWS has also been working on the infrastructure needed to make agents persistent and capable of managing their own execution loops.
In June, the company made the AgentCore managed harness generally available. The harness runs the agent orchestration loop, executes tools, manages context, preserves state across turns, recovers from failures and isolates individual sessions.
Developers can define an agent by specifying its model, tools, skills and instructions. AgentCore then provides the environment needed to run it, including filesystem and shell access, memory across sessions and web browsing.
AWS also added Web Search on AgentCore, giving agents access to current web information. AWS said the managed search system combines its web index with structured knowledge graph data and returns ranked results, relevant snippets, source URLs, titles and publication dates.
AgentCore Runtime also received interactive shell support in June. AWS said developers can open a persistent terminal inside a running agent session, inspect files, execute commands and reconnect to the same shell after a network interruption. A single agent runtime can support up to 10 concurrent shells.
Memory has been another part of the infrastructure. AgentCore Memory provides short-term context and long-term memory, while a June update added strictly consistent metadata for long-term memory records. AWS said the feature can help keep information separated by department, compliance boundary or tenant.
AWS is also adding controls over what agents are allowed to do. Its AgentCore Policy capability provides centralized rules for agent-tool interactions, with policies evaluated before requests reach tools. AWS says those policies can be authored in natural language and converted to Cedar, its open-source policy language.
The company has simultaneously been deploying agents for specific software and infrastructure tasks. In March, AWS announced general availability of its AWS Security Agent and AWS DevOps Agent, describing them as systems that can independently pursue goals, handle concurrent tasks and continue operating for hours or days without constant human oversight.
AWS Security Agent is designed for security testing, while AWS DevOps Agent is intended to investigate operational incidents and identify root causes. AWS described these systems as part of a broader category it calls “frontier agents.”
The company’s approach is therefore extending across several layers: compute for running agents, tools and web access, memory, identity, policy controls, monitoring, agent discovery and payments.
AWS is also connecting agents to existing workflow systems. Its June integration between AWS Step Functions and AgentCore allows agent reasoning steps to be inserted into workflows, including workflows that run multiple agents in sequence or in parallel and require human approval before critical actions.
In July, AWS said AgentCore had default capacity for thousands of concurrent sessions, while the newer Runtime and Runtime Instances are designed to support different patterns ranging from short-lived tasks to workloads that continue for days.
Taken together, the announcements represent a broader infrastructure effort around software that can reason through a task, use external tools, retain information, operate for extended periods and interact with other services.
AWS has not said that chatbots are being replaced. Instead, its product announcements show an expanding layer of infrastructure for applications in which the AI system can carry out multiple steps rather than simply generate a response to an individual prompt.
The September 18 Runtime release is the latest addition to that infrastructure. Its faster startup times and elastic memory system address the cost and performance of running agent workloads, while AWS’s other AgentCore services address persistence, tools, security, governance, discovery and transactions.
The direction is reflected in AWS’s own description of agents as systems capable of operating continuously across complex workflows. The company’s infrastructure is increasingly being designed around those requirements rather than around the shorter request-and-response pattern associated with traditional chatbot applications.
Discover more from Aree Blog
Subscribe now to keep reading and get access to the full archive.
