A Thought On LLM Pair Programming

2026-08-05

This thought was a part of a longer post which is mostly written, but unfortunately still mostly confusing and potentially unnecessary. One day that post may see the light, but until then I want this thought out in the open. It is not complete, nor necessarily right, but I believe it is worth considering.

Although LLM-powered development is still in its infancy, pair programming has been studied and tried to no end. In an agentic loop, a human participates in collaboration with the LLM as its navigator, analogously to a pair programming session.

By the word "participates", I imply that the programmer now contributes to something which is shared with something other than himself. When engineers pair program, both share some ownership of the code which is written. When programming with LLM assistance, the engineer shares some ownership with the LLM instead. A LLM not being a human who can take responsibility, part of the ownership has no shoulders to rest on.

In pair programming, the displaced ownership from one person lands on another. In LLM-powered development, the displaced ownership finds no home. The usual strategy for dealing with this homelessness: attempt to direct that ownership back to the programmer. While an honorable attempt at salvaging the situation, this leads on average to discontent. It relies on a myth: people are pleased to take responsibility for what they do not create. One may willingly set aside their displeasure, drawing on their better natures to take the blow in stride. Most won't.

One of the primary questions then for LLM usage is then: Who is held accountable for this displaced ownership? I wish I knew the answer, but I suspect LLM users won't like it in any case. The ownership debt must be paid somehow, and in the absence of willing sacrifice there will be exaction.