Amazon Bedrock Managed Agents (Preview): What to Check Before an AI Agent Pilot on AWS

Many companies want AI agents but hesitate over one question: where does the agent run, and who controls what it can do? On September 29, 2026, AWS announced a public preview of Amazon Bedrock Managed Agents, powered by OpenAI. The pitch is that you can run agents built on OpenAI models inside AWS, under the identity and governance controls you already use.

This guide explains what AWS has published, what remains unknown, and how to scope a first pilot. The service is a preview, so treat it as something to evaluate, not to depend on.

What Is Bedrock Managed Agents?

According to the AWS announcement, the service was developed jointly by AWS and OpenAI and is built on a customized version of OpenAI’s Agents API, made AWS-native. AWS says that it:

• Manages how the model keeps state, chooses and uses tools, runs code and coordinates multi-step work.
• Keeps durable sessions with messages, tool calls and intermediate results, so work can continue later.
• Lets you add reusable skills and connect tools, including through Model Context Protocol (MCP) servers.
• Gives each agent its own IAM role, supports human approval before consequential actions, and records supported API activity in AWS CloudTrail.

The AWS documentation says tools run in a compute environment that you provide: either your own host or Amazon Bedrock AgentCore Runtime. It also notes that preview features and APIs can change, and that deleting a session does not delete files you store in your own buckets or hosts.

What to Know Before You Plan

• It is a preview. AWS states preview functionality and APIs can change and points to preview limitations. Do not build a customer-facing service on it without a fallback.
• Regions are limited. AWS lists US East (N. Virginia), US West (Oregon) and US East (Ohio) for the preview. If your data must stay in another region, check this first.
• Pricing may change. AWS says pricing is subject to change at general availability. Resources your agents use, including some example stacks, can keep incurring charges. Review AWS’s pricing pages and clean up test resources.

How to Scope a First Pilot

These steps are our recommendations, not AWS guidance.

Pick One Narrow, Reversible Task

Choose a task with a clear owner and a clear success measure, such as drafting a summary of a support case for a person to review. Avoid tasks that spend money or change customer records on the first try.

Give the Agent the Least Access It Needs

Because each agent has its own IAM role, create a role that allows only the data and actions the pilot needs. Deny everything else. Review the role before each expansion.

Require Human Approval for Consequential Actions

Decide which actions a person must approve, such as sending a message, changing a record or running a command in production. Test that the approval step cannot be skipped.

Plan Logging and Review

Check which activity AWS records, and agree who reads the logs and how often. AWS says it records supported API activity, so confirm that your key actions are covered.

Measure Against a Baseline

Compare time saved, error rate and the share of cases a person had to correct against how the task works today. Stop or change the pilot if the numbers do not improve.

The Rule to Adopt

Give every AI agent, on any platform, only the access its task needs, require a person to approve consequential actions, and keep a log that someone actually reads. Make this the entry rule for every agent your company runs, before the first pilot, not after the first incident.

How Incresco Helps

Incresco builds web and mobile applications, cloud integrations, Salesforce solutions, WhatsApp experiences and AI agents. For an agent pilot on AWS, we would define the task and success measure, design roles and approvals, connect the agent to your systems and set up monitoring. This matches the approach on our AI transformation and cloud and DevOps pages: business outcomes first, then data and integration, then the technology choice.

Planning an AI agent pilot on AWS or another platform? Contact Incresco and ask for an AI agent pilot scoping session. We will help you pick one narrow task, define the access and approval rules, and set the measure that decides whether to continue.

Sources and Scope

• AWS, What’s New, September 29, 2026: preview announcement, capabilities, regions and pricing note.
• AWS documentation: Bedrock Managed Agents, powered by OpenAI (preview): execution environments, preview status, charges.
• AWS Weekly Roundup, October 5, 2026: context and related launches.

Sources checked October 8, 2026. We did not test the service. AWS descriptions are attributed to AWS. Regions, features and pricing may change.