AWS: Works Human Intelligence migrates from LangGraph to Amazon Bedrock AgentCore and achieves 97% cost reduction per process

Amazon Web Services published a case study showcasing the production use of Amazon Bedrock AgentCore — a managed AI agent platform — in collaboration with Works Human Intelligence (WHI), a Japanese HR software provider serving some of Japan’s largest corporations. The migration from LangGraph and Amazon ECS infrastructure to AgentCore resulted in dramatic cost reductions and simplified development.

What is Amazon Bedrock AgentCore and how does it differ from LangGraph?

Amazon Bedrock AgentCore is a managed service providing AgentCore Runtime as the execution foundation for AI agents, along with integrated observability and multi-agent workflow management. The key advantage over LangGraph + Amazon ECS is the elimination of manual infrastructure management — teams can focus on business logic rather than operations.

Works Human Intelligence previously used LangGraph (a framework for building agentic workflows) in combination with Amazon ECS (Elastic Container Service) to host their agents. This combination required manual log tracking and infrastructure administration. AgentCore Observability replaced the third-party Langfuse tool they used for monitoring, centralizing everything within the AWS ecosystem.

How does Works Human Intelligence implement the Commuting Allowance and Browser Operation agents?

WHI built two production agents:

The Commuting Allowance Agent automates commuting allowance approval processes for employee relocation. Slack serves as the entry point with authentication at invocation; sub-agents run on separate AgentCore Runtime instances, and multi-tenancy is managed through Amazon DynamoDB and Amazon Cognito.

The Browser Operation Agent accesses HR systems via browser automation, collecting records and performing operations. It is built using the Strands Agents framework, with AgentCore Runtime hosted within a VPC with a NAT gateway for a fixed IP address. Amazon S3 stores short-term operational data, and the knowledge base holds operation templates and support procedures.

How was the 97% per-process cost reduction achieved?

The Browser Operation Agent went through three staged cost optimizations:

1. Prompt caching implementation: $14.50 → $2.10 per process (85% reduction)
2. Agent behavior optimization: $2.10 → $1.00 per process (52% additional)
3. Model switch: Claude Sonnet 4.5 → Claude Haiku 4.5: $1.00 → $0.40 per process (60% additional)

Total: 97% cost reduction — from $14.50 to $0.40 per process. In addition to costs, the agent achieved an 88% reduction in token consumption in browser operations by removing unnecessary conversation history, eliminating irrelevant Playwright MCP return values, and caching tool definitions.

WHI evaluated three browser automation tools — browser-use, Playwright, and fast-playwright — with fast-playwright selected as the final solution due to the lowest token consumption.

Significance of AgentCore for AI production in enterprise systems

Amazon Bedrock AgentCore positions itself as AWS’s solution for lowering the barrier to production deployment of AI agents in enterprise environments. The Works Human Intelligence case demonstrates that complex HR processes requiring interaction with legacy systems (browser automation over HR applications), multi-tenancy, and secure authentication can be successfully automated with AI agents with measurable savings.

The WHI team notes that AgentCore “significantly simplified development” and enabled them to focus on business logic development rather than infrastructure operations — which is the central value proposition of managed agent platforms as a class.