Corporates make tasks idiot-proof by chopping the task into such modular pieces that each piece is really hard to mess up, and the next part of the process can pick that up and continue to do the same.

I think we should have a similar approach for agents and prompt chaining, that each message we send to an LLM carries a piece of work and a prompt that is really difficult to mess up, and it returns a result that is modular and can be used for the next process.

Basically, the error can stem from 3 parts.

• The prompt trying to do too much
• The result that the prompt is trying to create is not idiot proof.
• The result is not formatted properly, so the next prompt can’t pick it up.

The first one sucks because with a long, convoluted prompt, the model during inference can’t perform optimally due to cognitive overload. Second one results in the walk not getting completed properly because it was too complex. And third is that the next prompt, which is going to continue from where the previous one left off, It has an objective and way of working which is not compatible with where we left off. So it is a handoff problem.

To put simply, the problem is:

• The capacity of LLM itself to hold a lot of instructions and steps..

• The step is too complicated.

• Hand off between steps.

To fix this, we need to have a harness that first takes an input from you and then is able to break it into granular steps that take into account all three factors we just discussed.

We need a strong task decomposition layer between us and the LLM, and an execution layer that actually follows it.

Of course, you are still going to have detailed prompts and plans, but that is between you and the decomposition layer, not you and the model directly.