- Author
- Posted by emersonmacro, with replies from many commenters
- Published
- 19 January 2026
- Kind
- Hacker News discussion thread on a blog post
- Original
- news.ycombinator.com/item?id=46674416
- Subject
- Harnesses and agent loops
What it is
The linked article is "The Code-Only Agent" on rijnard.com. The thread has 155 points and was posted by emersonmacro. Replies compare one-tool agents with agents that have many specialised tools.
What it says
- Commenter binalpatel gives an agent one bash tool and has it write reusable CLIs.
- Binalpatel reports daily use, with generated CLIs for news, weather and a todo manager.
- Commenter rcarmo says the agent often builds a different tool in the next session.
- Commenter NitpickLawyer replies that notes in agents.md and skills already fixed that problem.
- Commenter killerstorm says models already know common tools like ls. Wrapping them distracts the model.
- Commenter znnajdla says one script is easier to revert and rerun than many calls.
- Commenter river_otter says a small set of flexible tools now beats the single-tool approach.
What evidence it gives
The thread offers anecdotes and opinion, with no controlled comparison. Several commenters link small open-source projects they wrote. One commenter links an arXiv paper on agents that write their own skills.
Interests and limits
Commenters are self-selected and many promote their own tools. The linked article was not read, so its claims reach this page only through commenters. Claims about cost and reliability are unmeasured. Some replies are jokes or speculation.