Summary
Mike Chambers, Senior AI Specialist Developer Advocate at AWS, presents at AI Engineer on what he calls harness engineering — the discipline of designing everything around a model that makes an agent production-ready. The core definitional move: take an agent, remove the model, and what remains is the harness. That reframe drives the talk’s practical guidance on how agents we build (deployed to users) differ fundamentally from agents we use (like Claude Code or Cursor), and why constraints appropriate for one category are wrong for the other.
Chambers distinguishes two agent archetypes, then moves into live coding using the Strands agents SDK inside the Kiro IDE. He walks through a minimal tool-equipped agent, then a more complete version incorporating a session manager for medium-term memory persistence — the agent can remember facts about the user across invocations by writing to files and rehydrating conversation history on the next session. The MCP tooling on display draws on his own MCP Lambda handler, which he notes still receives roughly 35,000 downloads per month since its 2025 release.
AWS’s founding membership in the Agentic Foundation (part of the Linux Foundation) gets a brief mention as context for where the industry’s shared standards effort is headed. The talk is aimed at engineers building and deploying domain-specific agents at scale, with emphasis on the structural decisions — memory architecture, session state, tool permissions — that separate a working demo from a production system.
📺 Source: AI Engineer · Published September 14, 2026
🏷️ Format: Keynote Launch
