An agent is made up of just two parts: the LLM and the harness. But what exactly is a harness? Is it the core “ReAct loop,” tool interface, and context management? Or the instructions, memory layer, skills, and context that you provide? Spoiler alert: the answer is “yes.” This talk is for anyone who wants more out of their agents, whether you're building agentic applications, using coding agents like Claude Code or Codex, or all of the above. More predictability and reliability. Higher quality output. How to be the engineer in the loop, not the agent babysitter. As many in the industry move toward “thin” harnesses, this talk argues for a strong “outer” harness: a coordination and control layer outside the core agent itself. Let the model reason and judge. Let the harness make the work repeatable, observable, and safe.
Paul Caplan is a Principal Engineer on the Developer Experience team at Teladoc Health, where he builds SDLC agents and helps drive adoption of AI tooling across the engineering organization. By night, he is building Codagent, an open-source harness toolkit for AI coding agents, and writing the “Agents, Harnessed” newsletter.
