Topic 3 of 9 · Level 2: Agent-assisted work
Harness Engineering
A harness is the suite of tools wrapped around a raw model. Codex, Cursor, and Claude Code are popular examples. I use Pi because it is open source, highly customisable, and cost efficient. It is CLI-based and more “batteries not included”; OpenCode is the open-source GUI or batteries-included counterpart.
A harness is personal infrastructure
With any model released in 2026 and a basic level of context engineering, a harness is largely a matter of personal preference. Some people prefer a GUI; some prefer a CLI. The right choice depends on how much control, setup, and automation you want.
The following is a personal ordering, not a universal ranking.
Harnesses I have used
Pi
This is the harness I use. It best suits people who enjoy CLI-based work and have a basic understanding of software engineering, although anyone can use it with a steeper learning curve.
Codex
OpenAI’s coding harness is mature, easy to use, and suitable for most people. In my experience, its cost efficiency is a little bloated.
Claude Code
Anthropic’s mature, easy-to-use coding harness was central early in my learning journey. In my experience, it is more bloated than Codex.
Hermes
It natively combines context engineering with a focus on self-improving agents. It can work well with enough use, but it did not meet my cost-optimisation objective.
Karimo
It offers a strong visual explanation of what a good harness can be, but I did not experience a material cost-saving difference.
Well-known options I have not meaningfully used
Why I stay with Pi
My primary objective is automation. Automation is useful only when outputs are known, traceable, and reliably good. Pi supports that style of work in a cost-effective way, especially because it lets a model write short TypeScript extensions where deterministic behaviour is needed.