• Home
• News

AWS Is Building a Cloud Where AI Agents Keep Working After Humans Log Off

The Chatbot Is No Longer the Right Mental Model

Cloud computing was originally designed around applications. A user opened software, submitted a request, a server processed it, and the cloud returned a result.

Agentic AI is changing that model.

An AI agent can begin working after an invoice arrives, continue processing a task for hours, call multiple tools, interact with another agent, wait for managerial approval and then resume from exactly where it stopped.

In some cases, there may be no human prompt at all. The event itself becomes the prompt.

That is the infrastructure challenge Amazon Web Services is increasingly addressing as it prepares for GITEX GLOBAL 2026 in Dubai.

AWS is moving beyond providing AI models through Amazon Bedrock toward creating an operating environment for autonomous AI systems through Amazon Bedrock AgentCore, managed agents, persistent runtimes and increasingly sophisticated connections between agents and enterprise infrastructure.

The shift is subtle but significant: the cloud was built to run applications, while AWS now wants it to run software-based workers.

The Innovation to Watch: Amazon Bedrock AgentCore

Amazon Bedrock AgentCore is AWS's effort to provide the infrastructure required to operate AI agents at scale.

Instead of requiring developers to build every operational component around an agent themselves, AgentCore provides services for running, connecting, securing, observing and managing agents in production.

Developers are also not locked into one particular agent framework or model. The infrastructure can support agents built using different frameworks and models.

That flexibility is central to AWS's strategy. The company does not necessarily need every enterprise to build the same AI agent. It wants those agents to run on AWS.

An AI Agent Needs More Than a Model

A language model can reason about a task, but that does not automatically make it a functioning enterprise agent.

An agent needs somewhere to run, an identity, memory in many cases, secure access to tools and APIs, monitoring capabilities and policies defining what it can access.

Security teams also need logs, while developers need ways to diagnose failures and understand agent behavior.

AgentCore separates these requirements into infrastructure services, reflecting a broader shift in the AI market.

As foundation models become increasingly interchangeable for certain workloads, enterprise competition is increasingly moving toward the infrastructure surrounding those models.

Runtime Turns an Agent Into a Persistent Workload

One of AgentCore's most important components is Runtime.

It provides managed computing infrastructure where agents can execute without developers having to build and maintain their own hosting environments.

AWS significantly updated the runtime in September. The platform is designed for agents that can operate for extended periods rather than completing a conventional short API request.

Runtime Instances can provide persistent volumes, GPU resources and multi-day sessions.

This capability matters because complex agents may need to maintain context while performing long-running work.

A cloud migration agent, for example, could inspect an application, generate infrastructure code, test assumptions, encounter an issue, request human input and continue later. The underlying computing environment needs to survive that workflow.

Some AI Agents Will Work for Days

This represents one of the biggest differences between agentic computing and conventional generative AI.

A chatbot session is generally temporary. An autonomous task may not be.

AWS has already demonstrated multi-agent workflows running on AgentCore Runtime Instances, with multiple agents sharing computing resources and a persistent filesystem while collaborating on longer tasks.

The same principle could extend to software development, scientific research, financial analysis, media production, infrastructure migration and security investigations.

In these scenarios, agents may operate continuously instead of simply producing a single response.

Cloud infrastructure therefore needs a new understanding of persistence. The workload is no longer just an application process. It can become a digital worker with unfinished business.

The Most Interesting Agent May Be the One Nobody Prompted

AWS is also developing what it calls ambient agents, changing how an AI task begins.

Traditional AI waits for a person to ask a question. An ambient agent responds to an event.

A document could appear in an Amazon S3 bucket. A security alert could be generated. A scheduled deadline could arrive. A business system could record a particular transaction.

The event activates the agent.

The agent can analyze what happened and automatically begin the appropriate workflow, making AI behave less like software waiting behind a chat window and more like an autonomous operational layer surrounding the business.

When the Event Becomes the Prompt

Consider an insurance company receiving a new claim document.

In a conventional workflow, the document might enter a queue until an employee opens it.

An ambient agent could begin immediately by identifying the document, extracting relevant information, comparing the claim with available records and identifying missing information.

If the case falls within its authorized workflow, the agent can continue. If human judgment is required, it can pause and request approval.

Once the employee responds, the agent resumes.

There was never a conventional prompt. The arrival of the document was the prompt.

That could become one of the most important transitions in enterprise AI.

Human Approval Does Not Have to End the Workflow

Autonomy does not necessarily mean removing humans from the process.

AWS's ambient-agent architecture includes human-in-the-loop mechanisms that allow an agent to pause when it reaches a decision requiring approval.

The human responds, and the workflow continues from its previous state.

This is particularly important for business processes that cannot legally, financially or operationally be delegated entirely to AI.

An AI system may be allowed to prepare a payment but not authorize it. It may analyze an insurance claim but require human approval for a large settlement. It could recommend security remediation while requiring an administrator's approval before modifying a production system.

Agentic infrastructure therefore needs to understand not only computing, but waiting.

AWS Is Bringing OpenAI Agents Inside Its Own Cloud Controls

One of AWS's significant recent moves came at the end of September, when the company introduced Amazon Bedrock Managed Agents powered by OpenAI in public preview.

The service uses a customized version of OpenAI's Agents API engineered to operate as an AWS-native environment.

The development creates an interesting relationship between two companies often discussed as participants in competing AI ecosystems.

Enterprises can build agents optimized for OpenAI models while running them inside AWS infrastructure and applying the identities, permissions and governance controls they already use.

The move demonstrates how AWS views the model market: it does not necessarily need to own the model to own the infrastructure underneath the agent.

Model Choice Is Becoming an AWS Weapon

Amazon Bedrock has consistently followed a multi-model strategy, providing customers with access to models from multiple AI developers instead of requiring them to use one model family.

The agentic era could make this strategy even more valuable.

A single agent may not require the same model for every task. One task could require advanced reasoning, another could prioritize speed, while another could favor a smaller and less expensive model.

An enterprise might also prefer one provider for coding and another for multimodal analysis.

Over time, agent infrastructure could select models according to the task, shifting strategic value from owning a single model toward orchestrating multiple models.

Agents Need Identities Too

Giving an AI agent access to enterprise systems creates an obvious security challenge.

An AI system should not receive unrestricted credentials simply because it needs to perform one task.

AgentCore therefore includes identity capabilities designed to control how agents access services and resources.

This reflects a broader transformation in enterprise technology.

Humans have identities. Applications have identities. Now agents need identities as well.

The question is no longer simply whether a model can perform an action. It is whether a particular agent is authorized to perform that specific action for a specific user.

Memory Is Becoming Infrastructure

Agents also need memory.

A useful digital worker cannot necessarily start from zero every time it receives a task. It may need to remember previous interactions, user preferences or information accumulated during an ongoing workflow.

But persistent memory creates its own questions.

What should an agent remember? How long should information remain available? Which agents should access it? Can memories cross user boundaries? How should sensitive information be protected?

AgentCore includes managed memory capabilities because memory is increasingly becoming an infrastructure concern rather than simply a feature of the language model.

Observability Becomes Critical When AI Makes Decisions

Traditional cloud systems produce logs and metrics. Agentic systems require deeper visibility.

Administrators may need to understand not only that an action occurred, but why the agent decided to perform it.

They may need to know which tool was called, which model was used, how long each stage took, where the workflow failed and what another agent contributed.

AgentCore Observability is designed to provide visibility into these workflows.

This becomes particularly important as multi-agent architectures expand. A final result may emerge from a chain involving several agents and external tools, making tracing essential for diagnosing incorrect results.

AWS Is Already Using Agents to Rebuild the Cloud Migration Process

One of the most revealing AgentCore examples comes from AWS Professional Services.

AWS has described a multi-agent architecture used in a migration program covering more than 300 applications.

Different agents handle tasks including discovery, infrastructure-as-code generation, governance and post-migration operations.

According to AWS's internal project tracking, the approach reduced infrastructure-as-code development from around three to four weeks per application to minutes.

The example highlights where agentic AI could create major enterprise value: not simply by producing marketing content, but by automating technical work that previously required weeks of specialist effort.

AWS Now Has an AI Agent That Inspects AWS

The recursion becomes even more interesting with the AWS Well-Architected Agent.

Introduced in public preview on October 1, the agent analyzes an organization's AWS environment and provides contextual recommendations around cost, security, performance and resilience.

It can also propose ready-to-implement fixes.

The Well-Architected Framework has traditionally helped cloud architects evaluate systems against AWS best practices. The agent begins automating part of that expertise.

This points toward a future in which cloud infrastructure can continuously examine itself rather than waiting for periodic architecture reviews.

Cloud Management Could Become Continuous AI Supervision

That model could significantly change cloud operations.

Today, engineers use dashboards to monitor infrastructure. In the future, agents could continuously monitor those dashboards on their behalf.

A cost agent could identify inefficient resources. A security agent could detect a risky configuration. A resilience agent could identify an architectural weakness, while a performance agent could locate a bottleneck.

Each could recommend action, with some potentially performing remediation automatically when given the appropriate authorization.

Cloud management could therefore move from human observation toward machine supervision with human escalation.

The UAE Gives AWS a Local AI Infrastructure Story

AWS's GITEX presence also has a direct regional infrastructure dimension.

In February, the company made Amazon EC2 G6e instances available in its Middle East UAE Region.

The instances use NVIDIA L40S Tensor Core GPUs and support workloads including large language models, image, video and audio generation, 3D simulations and digital twins.

Configurations support up to eight L40S GPUs with 48GB of GPU memory each, alongside up to 192 vCPUs, 400 Gbps of network bandwidth and 1.536TB of system memory.

For organizations in the Gulf, regional GPU availability matters because not every AI workload needs to travel to another continent for processing.

Sovereignty Is About More Than Where the Server Sits

The regional AI discussion is also increasingly focused on sovereignty.

A UAE-based AI company may need to ensure sensitive datasets, model assets and customer information remain within controlled infrastructure.

AWS recently highlighted Lisan AI, a UAE company specializing in Arabic natural-language processing.

The company consolidated previously fragmented multi-cloud AI workloads onto an AWS architecture designed around regional data residency, security and compliance.

AWS reports that the deployment supports more than 250,000 active users and reduced inference latency by 80%.

This demonstrates why sovereignty is becoming closely connected to AI architecture.

The question is not simply where a model runs. Organizations increasingly need control over where the data, model weights, logs and identities surrounding that model reside.

Why GITEX Is an Important Stage for AWS

AWS's Dubai audience is well suited to this transition.

Governments are developing sovereign AI strategies. Banks want agents but require strict controls. Telecom operators manage enormous operational workflows that could be automated. Energy companies operate complex infrastructure, while retailers and airlines manage large customer-service environments.

Startups also want to build AI products without having to construct the infrastructure underneath them.

For all of these organizations, the question is moving beyond whether generative AI works. The harder question is how to operate it reliably at scale.

What to Expect From AWS in Dubai

AWS is confirmed as a GITEX GLOBAL 2026 exhibitor at Hall S19, Stand H19-A10, while Tanuja Randery, AWS Vice President and Managing Director for EMEA, is among the executives listed by the event.

The company has not yet published a complete schedule of GITEX-specific product launches, so additional announcements should remain unconfirmed until AWS makes them.

Its recent roadmap nevertheless points toward several likely areas of focus.

Amazon Bedrock will remain central to AWS's generative AI platform, while AgentCore offers the more forward-looking story as enterprises move from experimenting with agents to operating them in production.

Managed Agents powered by OpenAI demonstrate AWS's willingness to accommodate competing model ecosystems inside its infrastructure.

Ambient agents show how AI can begin working in response to events rather than prompts.

The Well-Architected Agent demonstrates how AWS is applying agentic intelligence to cloud operations themselves.

Expanding GPU capacity in the UAE also provides a regional infrastructure foundation for these software developments.

The Cloud Is Becoming a Workplace for AI Agents

Cloud computing spent two decades waiting for humans and applications to send it requests.

Agentic AI changes that relationship.

Software can now observe events, start work, call tools, collaborate with other agents, pause for approval and continue long after the person who initiated the process has logged off.

AWS is betting that the next generation of cloud computing will not simply host AI models. It will become the workplace where autonomous AI agents live.