It's funny how consultants will try to co-opt terms.
No, the term does not come from Birgitta Boeckeler's article in martinfowler.com. It was coined by Mitchel Hashimoto months earlier in his 'My AI adoption journey' blog post [1], and immediately picked up by OpenAI [2] a week later.
Also worth noting that this text is not only entirely AI generated, but meant for other agents to read, not us humans.
The Agent/harness has no meaningful distinction. The backing model(s) can only read through a harness. Call it an agent, call it a tool, call it a library. They are all harnesses if they are backed by a model.
Broadly speaking,
- Copilot only reads AGENTS.md
- Codex reads AGENTS and CONTEXT
- Claude reads CLAUDE
This ai-literacy is noise, muddling the definitions and suggesting yet-another-md-file.
I think most people that use LLM coding tools enough independently derive most of the stuff in the article. For example my approach to garbage collection is to sample paths within the project and chunks of file content, then assemble context (similarity search results, git blame) for the agent to use to assess cruftiness and if found schedule cleanup.
It’s fascinating to live through a the emergence of a new technology and to see people trying to make sense of it as they go.
I personally think that a lot of the words used around AI, harness, SKILLS, agents, RAG.. are make up words or close to it, words that do not have profound semantics, even though people are trying to, often after the fact, make sense of them.
It’s just popular words that are different enough that people like to use them to claim a new knowledge, or to market a product.
But we could use AI tooling instead of harness in the abstract, and it would be better to use more precise terms for more concrete use cases, agent loop, CLI, IDE..
It was a fun game at a time where papers were competing for attention, but now that it has become a proven technology, I hope we can find more precise and meaningful words.
Maybe code owners? I'm not sure if it'll do exactly what you want; but you can say that "such and such" files must be edited, or reviewed, by specific users.
Then, if someone opens a PR who isn't one of those people listed, it requires them to be a reviewer.
We use it for .editorconfig; specifically, we don't want a newcomer to come in, not like a style rule, and slip in a change.
What even is this "AI Literacy framework"? When I see such overengineering I look at what the author has done. In this case, I find the author runs a consultancy on engineering "Habitats for Humans and AI" (https://www.russmiles.com/), listing a bunch of books he did not write (the author names are conveniently cropped out). This doesn't even belong on LinkedIn.
I'm not the author -- but I am currently fortunate to be sitting in a workshop he is teaching on the topic and figured HN might like (especially since he gives a lot of his content out for free/OSS).
As someone who does loads of AI-driven dev and governance, I'm finding there are a lot of great nuggets here. Between him (chaos engineering) and Kent Beck (extreme programming) I'n a kid in the candy store and wanted to share.
No, the term does not come from Birgitta Boeckeler's article in martinfowler.com. It was coined by Mitchel Hashimoto months earlier in his 'My AI adoption journey' blog post [1], and immediately picked up by OpenAI [2] a week later.
Also worth noting that this text is not only entirely AI generated, but meant for other agents to read, not us humans.
[1] https://mitchellh.com/writing/my-ai-adoption-journey
[2] https://openai.com/index/harness-engineering/
And here is the original author on a podcast talking about it: https://open.spotify.com/episode/4FxEdjXldNhoh67KYVmbDu
I also have a self-evolving architecture on it, but you have to be careful, especially if you're using multiple models (still WIP).
Broadly speaking,
- Copilot only reads AGENTS.md
- Codex reads AGENTS and CONTEXT
- Claude reads CLAUDE
This ai-literacy is noise, muddling the definitions and suggesting yet-another-md-file.
I don't know any harness that solves 'Garbage Collection' in the way described here
(most harnesses accelerate context pollution and code base drift via instructions they embed into system prompts, tool descriptions and skills)
The prose is definitely written by a Claude-like
I personally think that a lot of the words used around AI, harness, SKILLS, agents, RAG.. are make up words or close to it, words that do not have profound semantics, even though people are trying to, often after the fact, make sense of them.
It’s just popular words that are different enough that people like to use them to claim a new knowledge, or to market a product.
But we could use AI tooling instead of harness in the abstract, and it would be better to use more precise terms for more concrete use cases, agent loop, CLI, IDE..
It was a fun game at a time where papers were competing for attention, but now that it has become a proven technology, I hope we can find more precise and meaningful words.
https://walkinglabs.github.io/learn-harness-engineering/en/
Then, if someone opens a PR who isn't one of those people listed, it requires them to be a reviewer.
We use it for .editorconfig; specifically, we don't want a newcomer to come in, not like a style rule, and slip in a change.
The Harness is the Thing
https://news.ycombinator.com/item?id=49452346
As someone who does loads of AI-driven dev and governance, I'm finding there are a lot of great nuggets here. Between him (chaos engineering) and Kent Beck (extreme programming) I'n a kid in the candy store and wanted to share.