Same model, different agent

This is the observation the whole site rests on, and you can check it without leaving home. Take a model, put it behind two different products and ask them the same thing. The results do not look alike.

It is not that one reasons better. The reasoning comes from the model, and it is the same one. What changes is everything else: which tools it has in front of it, what it remembers of what it already did, when a human gets asked, what happens when a command fails, how much context fits before someone decides what to throw away.

Those decisions are not made by the model. They are made by the harness, and they are engineering decisions somebody wrote down somewhere.

Model and harness

Expand ⤢

The four decisions you notice most

• One Which tools it sees A model with a dedicated editing tool does not work like one asked to edit with terminal commands. The second can do it, but it makes more mistakes and leaves less trace of what it changed.
• Two What it remembers When the conversation runs long something has to go. What gets dropped, what gets summarised and what is kept whole determines whether the agent remembers the instruction you gave at the start.
• Three When it stops Asking before deleting a file, or not asking. Running inside a confined space, or on your whole machine. These are policies, and every product picks its own.
• Four What it does when something fails Retry, change strategy, warn. An agent that insists on the same broken command ten times and one that tries something else on the second attempt can have the same model inside.

Why this took so long to become visible

For two years the public conversation has been almost entirely about models. Which one scores higher, which one costs less, which one holds more context. Those are figures you can put in a table, and that makes them comfortable.

The layer above had no common name, no figures and was almost never public. You felt it in daily use and could not point at it.

DeepSeek publishing theirs under an MIT licence changes that part of the conversation, because for the first time there is a complete piece of that layer anyone can read end to end.

What this does not mean

The model still sets the ceiling. An excellent harness over a model that cannot program does not produce an agent that programs. The ability to reason is not made up for with engineering.

What a good harness does is bring the result closer to that ceiling instead of leaving it halfway. And what a bad one does is waste a good model.