Agentic engineering is flooding the hull of what it means to maintain software today. You might be one of those thinking about jumping ship under the impression that we are nearing the end, not only of needing to know how to write and read code, but of computer science more generally.
This course considers and cultivates the skills needed to keep on seafaring as a software engineer, even as the tides turn toward machines that can produce code at a speed and scope all but unimaginable before 2025. As a student in this course, you will conceptualize, plan, and build an agential harness—a tool analogous to Claude Code, Codex, or Opencode—plank by plank. We will cover topics such as the performance characteristics and capabilities of available LLMs, the cost structure of the hardware needed to generate tokens with them, and the tradeoffs in play in the architecture of an agential harness. We will study the scope and specialization of agential harnesses, their privacy model, their user experience, and we will touch on advanced features such as agent orchestration, planning structures, and version control.
Upon completion of the course and its assignments, students will gain:
Knowledge of the performance characteristics of modern LLM inference with respect to their use in agential harnesses.
Knowledge of the landscape of models provided by frontier labs, local and open source models, and other LLM offerings.
Hands-on experience using agential harnesses to write software in the form of a specialized agential harness.
Experience reasoning about the tradeoffs in performance, user experience, and privacy inherent in all software, whether agentially engineered or otherwise.
This course is taught by Lachlan Kermode, who may be contacted by email.