Overview
Mike Chambers argues that agent engineering begins with a distinction between agents people use and agents developers build for others. A coding assistant can rely on a local harness containing files, memory, skills, tools, MCP servers, and organizational standards. A production agent needs a substantially broader harness: it must manage the execution loop, multi-tenant isolation, scaling, payments, identity, context, memory, observability, and evaluation. Chambers defines the harness pragmatically as everything in an agent except the model, then demonstrates an escalating development path. He starts with a local Strands agent using a system prompt and tools, adds a session manager and file-backed memories, and finally moves those responsibilities into Amazon Bedrock AgentCore. The AgentCore CLI can generate Python or TypeScript agent code, expose it through HTTP, MCP, or AG-UI, provision short- and long-term memory, support local live development, and deploy through infrastructure as code. Its components can also be adopted independently, such as adding managed serverless memory to an existing production agent. Chambers further suggests that many agent use cases may need little custom agent code beyond a model, system prompt, and connected tools, making a configuration-defined harness sufficient in some cases. The central conclusion is that harness engineering is the discipline of making agents operable, secure, observable, and scalable for real users.
Sections
Professional and Technical Progression
Milestones that establish Chambers's background and lead into the production harness demonstration.
- Chambers collaborated with a colleague and Dr. Andrew Ng on a three-week generative AI with LLMs course that was approaching half a million enrollments.
- He created an MCP Lambda handler for simple serverless MCP serving, later reporting roughly 35,000 downloads per month.
- AWS became a founding member of the Agentic AI Foundation under the Linux Foundation, with Chambers contributing behind the scenes.
- The demonstration began with a minimal local Strands agent using a system prompt, framework-managed loop, calculator, and time tool.
- A second local agent added a session manager, conversation rehydration, and a tool for storing user memories.
- The final demonstration used the AgentCore CLI and runtime to scaffold, test, configure, and deploy an agent with managed infrastructure.
Harness Architecture and Implementation Details
Specific implementation patterns, runtime capabilities, and configuration choices discussed in the demonstration.
- The minimal Strands agent imports an agent abstraction and tool decorator, supplies a system prompt, and attaches tools such as a calculator and a custom time function.
- The memory-enabled example uses a session manager to rehydrate conversation history between invocations and stores some persistent memories in files.
- The AgentCore CLI can generate starter agents in Python or TypeScript or accept developer-provided code.
- Supported agent access modes mentioned in the talk include HTTP, MCP, and AG-UI.
- The platform is described as framework- and model-flexible; the demonstration uses the Strands Agents SDK and the default model labeled "Sonic 4.5" in the transcript.
- Managed short- and long-term memory are provisioned as cloud infrastructure, operate separately and asynchronously from the running agent, and remain connected to it.
- Adding the Amazon Bedrock AgentCore application integration enables scalable runtime execution and multi-tenant isolation without requiring the agent author to implement all tenancy logic.
- The local development command is "agentcore dev"; it launches a browser interface connected to the locally running agent and reflects code changes.
- The deployment workflow uses "agentcore deploy" and infrastructure as code to provision runtime, memory, and other selected components.
- A code-free harness can be represented by simple JSON containing the selected model and system prompt, then deployed through the same AgentCore workflow.
Tools, Frameworks, and References
Named resources mentioned as implementation aids or background reading.
- Agent Toolkit for AWS: a free GitHub toolkit that gives coding agents AWS deployment guidance and encourages infrastructure-as-code practices.
- Amazon Bedrock AgentCore: a composable AWS platform for agent runtime, memory, identity, observability, evaluation, and related production capabilities.
- Strands Agents SDK: an open-source, model-first framework used throughout the demonstrations to construct agents.
- AgentCore CLI: a command-line workflow for scaffolding, locally testing, configuring, and deploying agents.
- MCP: a protocol used to connect agents to external tools and documentation services and optionally expose deployed agents.
- AG-UI: an interface option mentioned for building interactive chat-agent experiences.
- Kiro: Chambers's IDE of choice for the presentation and the generator of one demonstrated agent implementation.
- LangChain article on agent harnesses: cited as an alternative explanation of the harness concept.
- MartinFowler.com article on coding-agent harnesses: cited as another reference for understanding harnesses.
- Generative AI with LLMs course: a three-week course created with Chambers's colleague and Dr. Andrew Ng.