Mike Chambers is a Brisbane-based Senior Developer Advocate for Generative AI at Amazon Web Services whose work focuses on turning language models into reliable, deployable agents. His projects and teaching address the engineering problems that begin when models interact with external tools: preserving context, managing asynchronous work, securing infrastructure, and operating systems in production.
Chambers previously worked in solutions and security architecture across enterprises and government before becoming an independent cloud and machine-learning trainer. AWS recognized him as an AWS Machine Learning Hero in 2020; he subsequently joined the company and shifted his focus from machine learning to generative AI. He helped develop and teaches Generative AI with Large Language Models, produced with DeepLearning.AI and AWS, and teaches serverless agentic workflows with Amazon Bedrock.
Production agents need working memory. Chambers defines an agent through five practical components: a model, instructions, a reasoning loop, conversational history, and tools. Crucially, history must persist between tool calls within the same task, allowing each step to build on earlier results. His production-agent demonstration develops this architecture from a local dice-rolling program into an Amazon Bedrock application backed by AWS Lambda; he separately recommends infrastructure as code for production provisioning.
Serverless Model Context Protocol infrastructure. His Lambda-MCP-Server exposes agent tools over HTTP without maintaining a continuously running server; its handler subsequently moved into the AWS Labs MCP project.
Truly asynchronous agent tools. The async-agentic-tools project returns immediately when background operations begin, keeps the conversation responsive, and delivers completed results through callbacks. The approach is particularly relevant to voice assistants, where blocking on slow tools creates awkward silence.
Infrastructure OpSlop. Chambers uses this term for agents that create cloud resources through unreviewed, imperative actions without reproducible definitions. His case for agent-generated infrastructure as code advocates version-controlled templates, reviewed changes, limited permissions, and clear deployment boundaries; related work explores sandboxed tool orchestration without unrestricted access to networks, files, or secrets.
Build on software engineering skills by learning model behavior, using AI development tools, testing models through a common API, and composing tools into agent workflows.
A tabletop RPG assistant shows what must move beyond the laptop: model access, instructions, the agent loop, execution history, and the tools that act on its decisions.
Mike Chambers builds from a small Strands agent to persistent sessions, separately managed memory and a cloud runtime, then shows how a built-in harness can replace application code with configuration.
A harness is everything left after removing the model from an agent. Serving users adds runtime, identity, scaling, context management, observability and evaluation to the assistant's memory, skills and tools.
Session restoration carries conversation history across invocations; an explicit remembering tool retains selected information separately. The World Cup example shows remembered preference influencing a later answer.
Infrastructure as code keeps deployment definitions under developer control, whether an assistant helps create cloud resources or the AgentCore CLI deploys the agent and its surrounding services.
Separately managed memory changes the architecture: persistence becomes connected infrastructure outside agent execution, with its own scaling consideration.
▶ Watch ↗
