OpenHands

Are there any papers evaluating opinionated coding strategies for agents, and which ones tend to be more effective for reliability/easier to review for humans?

e.g include
• issue description + freely code
• issue description + TDD via single agent
• issue description + TDD via multi-agents (one writes test, other writes app code)
• etc

That's a good question! I do sometimes tell my agent to red/green.

Is this really about a coding strategy or about human understanding?

I don't have an answer, I do want to point out quickly a few examples which, to my knowledge, are/were experiments to get human understanding to where we need it

• a thing we could maybe try is to tell the agent to put together a clear HTML page with the component(s) after the PR; or the changes brought by the PR; or the flow of a feature. 
    ◦ I try to make a habit of it, and see where it leads, e.g.
    ◦ <https://enyst.github.io/arch/openhands-deferred-init-runtime-boundary.html|Deferred init runtime boundary>
• (better) Xingyao has been working on a `show me` skill: <https://github.com/xingyaoww/show-me|github.com/xingyaoww/show-me>