Why One Giant AI Agent Is the Wrong Abstraction For a Harness

For long running agentic work, picking a super capable model is a waste of money most of the time.

Even with super capable models there is still a context limit where the model falls off in terms of quality. Keeping contexts small is paramaunt to getting more reliable work done. That is why instead of thinking of more specialised instructions or more well-defined skills or prompts, you should think of state transitions instead.

One agent vs State transitions

Shifting to state transitions allows us to think about a workflow in smaller goals. The overarching goal is still there, what we are trying to fix, but the destination of the steps are different.

The code execution step is to open a PR.

The CI check step is to make the PR green.

The merge step is to get things to production without a risk.

Smaller goals, less context preload, more focused execution, and as a consequence, less money spent.

The additional benefit of state transitions:

  • Easier to debug and evaluate

  • Easier to extend

  • Human in the loop opportunities