Don't let them make you forget about software.
There's no AI agent, nor any harness. These are just buzz words coined to take the emphasis away from software and software development.
There's just software, machine learning models made using software, and software that makes good use of those models.
If you divide the digital world into:
Software: deterministic programming + non-tensor data e.g. control logic, user inputs, databases, APIs, heuristics etc.
Machine Learning Models
Then, every digital AI system that exists today completely falls in either of those two categories, often both.
There is no mystical third element!
Everything running on servers (CPU / GPU) consists entirely of traditional software, and tensor-based models. Nothing more.
The funny part is: most users never get even close to those models. All they use are software that use those models, and then call it Agent, Harness, etc. etc.
Models don't even do any tool calls either! They only accept input (text, image, audio, video) and return output (text, image, audio, video).
That's all there is to it. It's the software that makes the tool calls based on what the models return.
The models don't have any control over what happens after it returns something.
We do it using the software we developed, software uses those models, then the users do it based on those software.
Once you understand this, use those buzz words when you need to, but every once in a while, remind people of what they are actually using under the hood.
P.S. of course, most developers are already aware of the under the hood architecture of the whole thing. However, what most people are probably not aware of, is the fact that, even switching the word software with terms like "agent" and "harness" devalues our work to people who don't know any better. That's why, saying out loud what it is, at least some of the times, is important. 🥰
Top comments (3)
Right.
Left
🤔
Hi