fbeeper

If you’ve been building features with LLMs over the last few years you’ve likely used frameworks like LangChain/LangGraph, Vercel AI SDK, LlamaIndex, Google ADK, etc. Alternatively, you may be deeply exposed to chatbots, coding agents, and/or assistants. All of the above, roughly circling around the same idea, are flavors of the AI Agent loop being a mix of large language models and software scaffolding (typically called a harness).

What’s surprising is that for Swift and Apple platform developers, a solid version of the harness-building toolset doesn’t seem to exist yet. Working on some exploratory features, I kept having to relearn and rebuild the same scaffolding every time I wanted to try a different provider. The kind of thing a framework exists to solve. And since a clean Swift abstraction of it didn’t already exist, I built mine.

While I flattened the repository history before making it public, the repository history of AgentKitten spanned more than 570 commits and 130 pull requests of rethinking the shape of things. And I've continued to maintain it since, having piled up 64 more PRs to improve it further. It is built using AI coding tools, but steered carefully and deliberately.

I believe we still have a world of opportunities to put AI (particularly small on device models) to the service of improving mundane interactions in our apps (I put some samples in the repo’s README.md). I think this focus was kind of instrumental on how I built AgentKitten. It is more powerful that I expected it to be when I started it, but still I tried to stay grounded about being practical for the little use-case.

Clone it and run `swift run Playground --help` to see examples of what it can do.

Without widespread adoption is still an early stage project and I'm looking for constructive feedback to improve it.

My thoughts on the use of AI and its relationship with human reviews. A little toy sample sprinkled on the post to try illustrate the point.

For those non familiar to Ook! http://www.dangermouse.net/esoteric/ook.html

That somehow reinforces that publishing source code with CS research papers is more than necessary :)

I don't think that they even expect them to be predictors of importance or popularity. The "best paper" award can give a hint of temporary fashions or about what seemed to be interesting concepts/works at a certain point in time. That's why it seems especially interesting to see them compared with the "top cited" in these charts.