A few months ago, I became interested in Harness Engineering and started researching it. I realized that, in my experience, as tasks start to grow, agents end up losing context, repeating steps, or trying to solve the same problem multiple times. I wanted to build this workflow layer to provide a clearer structure for working on features, bug fixes, and patches, while persisting task state. Of course, Hanesu can make mistakes, but having a record of what was done helps a great deal in making it more likely that a second iteration will achieve the expected result.
So why not just use CLAUDE.md and put all the context there? Because CLAUDE.md describes how to work, but it doesn't save the execution state of a task. Context is limited, and as a conversation grows, it becomes more difficult to maintain the entire state of a task solely within the chat.
In addition, by using multiple agents, each one focuses on a specific stage of the workflow. Hanesu leverages existing tools (Codex, Claude Code, or OpenCode) to serve as a thin layer of control within the repository. Finally, this is still an experimental version; for small tasks, it’s best to use a traditional prompt, because it consumes more tokens due to the additional stages, so it is better suited for complex tasks.
I'd like to know if this approach makes sense to you; in my experience, at least, I've found it useful. Thanks for reading. :)
Comments URL: [https://news.ycombinator.com/item?id=49017839](https://news.ycombinator.com/item?id=49017839)
Points: 1