Hey {{first name | there}}. After spending a long week at the AI Engineer World's Fair in San Francisco, I got to thinking about how good the models have become.
More importantly, I started thinking about harnesses and whether they will end up making a bigger difference than the models themselves.
In today's Technical Notes:
What is a harness?
What does this have to do with anything?
Why I think the answer lies in the harness
📰 TECHNICAL NOTES: What is a harness?
Borrowing from Databricks:
An AI agent harness is the software infrastructure that wraps around a large language model (LLM) and enables it to act on tasks, not just respond to prompts. The model reasons through a problem and decides what to do next. The harness connects it to the tools, systems, memory, and execution environments needed to carry out those actions.
Using that definition, you can start to think of an agent as:
Agent = Model + Harness
The model is the brain. The harness provides the tooling. Following the human anatomy analogy, the harness is everything else. The arms, the legs, and the ability to interact with the world.
So what does this have to do with anything?
To some extent, every new model release improves on the previous one. Some models are better at coding. Others perform better at web search or reasoning.
The point I'm trying to make is that there now seems to be a baseline level of capability that most frontier models share before you even factor in their specialties.
The question then becomes:
How do you get more out of a model when they can all do roughly the same things?
Why I think the answer lies in the harness
I think the answer lies in the harness.
While I haven't used it in my day-to-day work yet, I think OpenCode is a fantastic example of what it means to build a better harness.
You bring whichever model you prefer, whether that's Claude, GPT, Gemini, or something else, and OpenCode provides a consistent developer experience on top. The value isn't just the model. It's everything surrounding the model.
That is what makes me think harnesses will become increasingly important.
As models continue to improve, the differences between them will matter less for many use cases. What will matter more is how well they are connected to tools, memory, execution environments, and the workflows people use every day.
🌍IN THE ECOSYSTEM
Together AI: One of the more interesting AI infrastructure companies I came across. While most conversations focus on foundation models, Together has been investing heavily in the serving layer.
OpenComputer: This project came up multiple times throughout the conference. The idea is simple but compelling. Instead of giving agents short-lived sandboxes, OpenComputer gives them persistent, long-running virtual machines. It feels like a strong direction for background agents that need state, memory, and the ability to run for hours rather than minutes.
AI Harness Engineering (Databricks): This article inspired how I've started thinking about harnesses. It argues that model capability is only one part of the equation. The software surrounding the model, including context management, tool access, memory, permissions, and verification, plays an equally important role in determining how useful an AI agent becomes. It's worth reading if you want to understand why I think harnesses are emerging as their own engineering discipline.
⏱️UNTIL NEXT TIME
Foundation models will continue to improve. Every few months we'll see another benchmark, another state-of-the-art release, and another wave of excitement.
What I'm watching is everything around the model.
The products that people use every day won't be defined solely by which model they run. They'll be defined by how well that model is connected to tools, memory, execution environments, and the workflows users already have.
That's why I think the next wave of innovation won't just come from better models. It will come from better harnesses.
Know an engineer wrestling with this? Share this link with them
Jubril Oyetunji
CTO, EverythingDevOps
