Building AI agents for business support using Amazon Bedrock AgentCore

Developing AI agents for business support presents unique challenges that many organizations face when trying to perform routine HR tasks. Works Human Intelligence (WHI) develops, sells, and supports an integrated HR system “COMPANY” for Japan's largest corporations and public interest companies.

In this post, we share how AWS Generative AI Innovation Center (GenAIIC) collaborated with Works Human Intelligence (WHI) to build two AI agents using Amazon Bedrock AgentCore. We discuss challenges encountered and solutions that reduced costs by 97 percent while improving efficiency.

Clients using HR systems must respond to many situations, such as organizational changes, HR system updates, and employee information updates. For organizations facing similar challenges with HR system performance, AI agents can significantly reduce workload and improve productivity. When WHI started building products using AI agents, several challenges arose. To solve these problems, we at GenAIIC worked closely with the WHI team to provide new ideas and support in creating a high quality product. The scope of this project includes two AI agents designed to support the work of operational departments. The Travel Permit Agent handles the approval of travel permit applications that arise during events such as movements. Browser Operating Agent “COMPANY” on behalf of the customer. We discuss the challenges and solutions for these two agents in the following sections.

Travel clearance agent

This agent automates the approval of travel permit applications, a common task that occurs during events such as employee relocations.

The challenge

The Travel Authorization Agent supports the routine task of approving travel authorization applications. WHI was already conducting a proof of concept (PoC) using LangGraph, Amazon Elastic Container Service (Amazon ECS), and AWS Fargate. However, because Amazon Bedrock AgentCore was released during development, the team began to consider migration. WHI wanted to work with us to build a solution with AgentCore that would see an AI agent integrated with a “COMPANY”. Additionally, they wanted to move to an integrated multi-agent environment and implement authentication and authorization using AWS Fargate and Amazon Cognito, which were currently under development.

Solution overview

The Commuting Allowance Agent was being developed using LangGraph and Amazon ECS, but the team had concerns about a monolithic configuration where everything ran as a single Amazon ECS task. Therefore, we worked together to change the architecture so that the sub-agents are launched individually in the AgentCore Runtime. Because multi-tenancy support was required, we decided to manage tenants using Amazon DynamoDB and Amazon Cognito to maintain WHI's flexibility to build and manage it.

Buildings

Slack serves as the entry point for calling the Travel Authorization Agent, so the system is designed to authenticate at the time of the call, and then subordinate agents process the request.

Results and impact

Since AgentCore is generally available (GA) during the project, we were able to use it successfully. While the Workflow Agent continues to use LangGraph, we've changed it so that the sub-agents run on a separate Runtime. This change facilitates the future expansion of sub-agents. We are also thinking of changing the director agent that includes the sub-agents to Strands Agents in the future. Additionally, while WHI used to manage Langfuse to check the status of agents, which incurs operational costs, switching to AgentCore Observability has reduced this burden.

Browser performance agent

This agent uses a browser to access the HR system, check content, perform tasks, and collect evidence on behalf of clients.

The challenge

The second agent uses a browser to access “COMPANY”, inspects content, performs operations, and obtains credentials. Development was ongoing with LangGraph and the Playwright Model Context Protocol (MCP). The team ensured an 88% reduction in browser performance tokens with the following methods:

• Removing unnecessary parts past the agent loop (dialog history between AI and Playwright MCP).
• Removing unnecessary parts of browser functionality from Playwright MCP return values.
• Using temporary storage of the INTRODUCTORY component.

However, because it relied on proprietary implementations, the team faced challenges such as the difficulty of migrating to Strands Agents. They were also thinking of ways to further reduce the tokens. It was in this context that GenAIIC began working with the WHI.

Solution overview

We built an agent using Strands Agents. After testing several browser performance tools and ensuring efficient processing, we focused on reducing the number of tokens used. The workflow starts by searching for the correct working template from the knowledge base according to the user's instructions. Next, it replaces the proxies in the received template with information obtained from another knowledge base to create a work manual. The agent then uses a browser based on this manual to check the current information. Based on the received information (such as CSV), it creates a change proposal and presents it to the user. Finally, after the user's authorization, we also use the browser based on the change topic to make the changes. Although a basic workflow is defined, the agent can handle situations where user input is insufficient or when only certain tasks are performed, based on its own decision.

Buildings

Access to “COMPANY” from the agent is limited by IP address. To address this, we placed the AgentCore Runtime inside a virtual private cloud (VPC) and configured it to access via a fixed IP address using a NAT gateway. We also built a knowledge base to store work templates and help information for creating work processes. We used an Amazon Simple Storage Service (Amazon S3) bucket to store short-term data.

Results and impact

The Browser Operation Agent is built using Strands Agents. We've tested browser performance tools including browser implementation, Playwright, and Quick Playwright, ensuring that Quick Playwright consumes very few tokens. In addition, by collaborating on improvements such as using Amazon Bedrock prompt caching and system prompts, we succeeded in reducing the cost per process by up to 97%. The main improvement measures were as follows:

• Using caching of user messages: Enabled caching feature of Amazon Bedrock ($14.5 -> $2.1).
• Improving agent behavior: Improved low agent notifications to reduce unnecessary performance ($2.1 -> $1.0).
• Changing models: Changed model from Claude Sonnet 4.5 to Haiku 4.5 ($1.0 -> $0.4).

With these improvements, we have improved costs while effectively managing complex operations. This includes situations that make multiple changes in a row or situations where an agent asks a person questions when the instructions are unclear.

The conclusion

Through our collaboration, we successfully moved the AI ​​agent implementation infrastructure to AgentCore Runtime and now we can check the performance status using AgentCore Observability. WHI members say that using AgentCore has greatly simplified development, as log checking is now done through a managed service. In addition, adopting Strands Agents for the Browser Agent allowed us to see an agent that behaves dynamically with minimal implementation. In this post, we explained how building AI agents can support common tasks. Our work together has allowed WHI to reach a position where it can focus on developing business intelligence. AgentCore includes Runtime, which serves as an execution base, and other features, so we consider future use with WHI. Additionally, the behavior and costs of AI agents change with the model used. With this project, we have ensured that processes work as expected. We plan to continue testing models and increasing costs.

To see for yourself how Amazon Bedrock AgentCore simplifies AI agent development, visit our Getting Started Guide or Hands-on Lab. Whether you want to automate common tasks, build multi-agent workflows, or improve model costs using features like caching, AWS has the tools to support you.

About the writers