AWS Introduces 'Harness Engineering' for AI Agents

Chambers has a background that includes working with Dr. Andrew Ng on a generative AI course focusing on large language models. He is also known for creating an MCP Lambda handler, which achieved approximately 35,000 monthly downloads in 2025, and highlighted AWS's role as a founding member of the Entka Foundation.

Mike Chambers identified two distinct categories of AI agents: those that users engage with directly and those that engineers construct behind the scenes. Agents users interact with include productivity tools such as Claude Code, Cursor, and Kirao, which enhance daily tasks. In contrast, agents that engineers build are complex systems that demand careful architectural design, robust infrastructure, and consideration of the target audience. While these two types of agents are conceptually separate, they often chain together in practical applications, where user-facing agents leverage the capabilities of expertly built back-end systems.

In this context, the 'animal' is replaced by an 'AI model,' meaning a harness is a set of control mechanisms applied to an AI model. Technically, Chambers specifies that an agent, when stripped of its core model, leaves behind the harness, which encompasses all the surrounding infrastructure and control logic.

This includes critical components like memory management, which ensures the agent can retain and utilize information across interactions. It also incorporates predefined skills and toolsets, enabling the agent to perform specific actions, and utilizes MCP (Model Context Protocol) servers for connectivity and interaction.

These new standards build upon the long-standing tradition of coding standards that have guided software development for decades. The focus is on defining consistent practices for deploying AI-driven tools to ensure reliability, scalability, and security across an organization's development environment.

This toolkit is designed to assist engineers in deploying code and managing AI agents effectively within the AWS ecosystem. Its availability aims to streamline the development process for those building and deploying agent-based applications.

This approach ensures that developers maintain clear ownership and control over their cloud deployments, rather than relying on manual processes for tasks like spinning up S3 buckets or EC2 instances. By leveraging IaC, deployments become repeatable, auditable, and scalable, aligning with best practices for cloud development.

Mike Chambers highlighted the need to manage the agent's loop logic, as well as scaling payments, memory, identity, and skills effectively. Runtime context management and robust observability and evaluation mechanisms are also essential for production-grade systems. Chambers stressed that bundling all these components into a single container is insufficient; instead, each must be independently scalable and manageable to ensure a resilient and high-performing system at cloud scale.

Mike Chambers demonstrated a simple agent built using the Strands Agents SDK. This agent was configured with a tool decorator, system prompts, and basic tools, including a calculator and a time retrieval function. While this setup is sufficient for local development and testing, Chambers noted that it lacks the essential attributes required for production-scale deployment, such as robust session management and scalability features.

Mike Chambers emphasized the critical role of a 'session manager' for scaling AI agents and handling complex, multi-turn interactions. He highlighted its inclusion within the agent's core design. A session manager allows the agent to track and maintain context across multiple interactions, ensuring that conversations remain coherent and relevant over time. This capability is vital for providing a seamless user experience and enabling the agent to perform more sophisticated tasks that require memory of past exchanges.

This framework simplifies the process by providing built-in capabilities for multi-tenant isolation and runtime scalability, significantly reducing the need for extensive custom code. By leveraging Strands, developers can efficiently manage critical aspects like security and identity, allowing them to focus on the agent's core functionality rather than complex infrastructure concerns.

AWS's 'agentcore dev' command enables developers to spin up a web browser interface for live local testing of their agents. This feature reflects code updates in real-time, allowing for immediate feedback during the development process. For production deployments, 'agentcore deploy' handles the creation of infrastructure as code, automating the process for scalable and repeatable environments.

Developers can deploy agents through a simple JSON configuration that specifies the desired model and system prompt, eliminating the need to write complex agent code for common use cases. This method effectively addresses approximately 80% of typical agentic requirements, allowing for rapid deployment and iteration.

This means individual components, such as serverless long-term memory management, can be adopted and integrated into existing production agents without requiring a complete overhaul. The design encourages modularity, providing flexibility for developers to pick and choose the features they need, rather than implementing monolithic solutions for custom agent requirements.