AWS Moves AI Agents Beyond Demos: Up to 14-Day Sessions, Controls, and Robots
AI agent
Software that follows instructions and takes several steps toward a task.
Amazon Bedrock
An AWS service for using different AI models.
AgentCore
AWS tools that help AI agents run and stay within limits.
Amazon Web Services (AWS), Amazon's cloud computing business, published a roundup on September 9. It covered AI builder updates released during August 2026. The main story was not one new model. It was the growing system around AI agents, software that can follow instructions and complete several steps.
What happened
Amazon Bedrock, AWS's service for using different AI models, added large context windows for OpenAI's GPT-5.6 Sol, Terra, and Luna. The window can hold one million tokens, or very small pieces of text. Prompt caching can reuse repeated information. Bedrock Web Search can retrieve fresh public information and return cited answers. Cross-Region inference can route requests across more than 25 AWS Regions. Developers can choose broad global routing or a geographic boundary.
Other updates track usage and spending. IAM principal cost allocation can connect inference costs to a person, team, project, application, or cost center. AWS Cost Anomaly Detection can flag unusual changes in third-party model spending.
Background: Agents Need More Than Good Answers
An ordinary chatbot often handles one request. An AI agent can search, call tools, remember information, and act through a workflow. That creates a different engineering problem. Teams must decide what data an agent can see, what actions it can take, and when a person must approve. They must also track time and cost. AWS's roundup shows attention moving from model ability to the surrounding system.
Why it matters
AgentCore runtime can run an agent on dedicated Amazon EC2 compute for sessions lasting up to 14 days. AWS names research, coding, and monitoring as examples. Temporal policies can judge an action using what the agent did earlier. Rate limits can restrict requests, tokens, and connections. AgentCore Payments can give agents access to paid APIs and resources under spending limits. AgentCore Memory can build long-term memory from business events. Fine-grained controls can separate that memory by user or tenant. AWS Agent Registry offers a searchable catalog of approved agents, tools, skills, and custom resources.
This changes the product question. It becomes less about whether a model can answer. It becomes more about whether an organization can delegate work with clear limits. That is an interpretation of AWS's update pattern. It is not proof that agents are reliable or safe.
What is confirmed
For regulated environments, AWS says Claude Opus 5 and OpenAI GPT-5.6 Terra and Luna are available in AWS GovCloud Regions. Claude Opus 5 has zero data retention enabled by default there. Nova Multimodal Embeddings supports retrieval across text, documents, images, video, and audio in AWS GovCloud (US-West).
On the physical side, Strands Robots connects Strands Agents, LeRobot, and Hugging Face Storage Buckets. AWS says builders can record demonstrations, stream data, train policies, and deploy from simulation to physical hardware without converting the underlying data. It also describes mesh discovery and coordination for multiple devices, using Zenoh locally and AWS IoT Core for distributed fleets.
AWS also says Bedrock has more than 225,000 active customers, including more than 80% of Fortune 100 companies. That is a company-reported figure.
What remains unknown
The post is an official product roundup, not an independent evaluation. A 14-day session limit does not mean an agent can work correctly for two weeks. AWS does not provide common error rates, completion rates, real customer savings, or the share of wrong actions stopped by controls. Access, pricing, data handling, and regional availability may differ by model and customer. The robotics section describes a connected workflow, but it does not show how safely or reliably it works in real settings.
What to watch next
The next test is real use. Watch whether long-running agents finish useful tasks without creating more supervision work. Watch how companies manage memory, permissions, payments, and emergency stops. Watch whether geographic controls meet regulated customers' needs. For robots, independent tests should show the gap between simulated practice and physical deployment.
Read the AWS official August roundup for the complete feature list.
AWS Gives AI Helpers Longer Jobs and More Rules
📰 Full story: AWS Moves AI Agents Beyond Demos: Up to 14-Day Sessions, Controls, and Robots
AWS is building tools that help AI agents handle longer tasks.
AI agent
Software that follows steps to complete a job.
AgentCore
AWS tools that help AI agents run longer and follow rules.
prompt caching
Saving repeated information so an AI can reuse it.
💡 The gist
- AWS summarized AI tools released during August 2026.
- The tools help AI agents work longer and use more information.
- They add rules for memory, money, regions, and robots.
Amazon Web Services (AWS) is Amazon's cloud computing business. It shared the roundup on September 9. Amazon Bedrock is an AWS service for using different AI models. AWS says Bedrock has more than 225,000 active customers. This number comes from AWS itself.
An AI agent is software that follows steps for a person. A chatbot may answer one question. An agent can search, use tools, and continue working. Longer jobs need stronger rules. People must control the data, actions, time, and spending.
Several August updates support longer jobs. Bedrock models can handle a very large amount of text. One new window holds one million tokens. Tokens are small pieces of text. Prompt caching can reuse information that appears again. Web Search can find newer public information and add citations.
AgentCore is a group of AWS tools for running AI agents. Some sessions can run for up to 14 days. Fourteen days means two weeks. Other tools check actions against earlier actions. They limit requests and spending. Agents can also use paid services with set limits. Memory tools can keep useful information separate for each user or organization.
AWS added a catalog for approved agents and tools. It also made some models available in AWS GovCloud. GovCloud is AWS's cloud for certain regulated work. Another project connects AI agents with robots. Robots can practice in simulations. Then builders can move their work toward real machines.
The roundup describes available features. It does not prove that every agent works well. It gives no common error rate or success rate. It also does not show real savings for customers. A long-running agent still needs human rules and checks.
The important change is the kind of work people may delegate. AI may handle longer chains of tasks. People must decide where the chain stops. They must also watch memory, money, and physical actions.
Read the AWS official August roundup.
AWS Helps Computer Helpers Work Longer
📰 Full story: AWS Moves AI Agents Beyond Demos: Up to 14-Day Sessions, Controls, and Robots
AWS says new tools can help computer helpers work longer.
AI agent
A computer helper that follows steps for a job.
Amazon Bedrock
An AWS place where people can use different AIs.
AgentCore
An AWS tool that helps a computer helper work longer.
Amazon Web Services (AWS) is a company that rents computer power. It shared news about its August tools. Amazon Bedrock is an AWS place for using different AIs. An AI agent is a computer helper that follows steps for you.
Before, a helper might answer one small question. New tools let a helper keep working longer. Some AgentCore sessions can last up to 14 days. Fourteen days means two weeks. The helper can remember useful things. It can check rules before taking actions. It can also have a spending limit.
Another tool gives helpers a shelf of allowed tools. AWS also showed ways to connect helpers with robots. A robot can practice in a pretend place first. Then it can try the same work on real machines. This does not mean every robot can do every job.
AWS says Bedrock has more than 225,000 active customers. AWS is reporting this number. The roundup does not prove that every helper works perfectly. It also does not show how often helpers make mistakes.
So the big change is simple. People may give computers longer jobs. People still need to set rules and watch the work.
Read the AWS official roundup.