Does it make sense to kind off simulate an agent mode

[vivek100]

I was thinking its slightly painful to do the changes one by one, thought its really amazing to speed up the copy paste and diff using chatgpt.

Do you think it makes sense to have agent mode which creates a plan, and then initiates chat one by one for the changes?

I can kind of simulate it by copy pasting the ai response back in the chat window in the extensions and ask it to do the next task.

[robertpiosik]

With automated task execution, context is impossible to be managed. This obstacle makes agents impossible to function reliably.

The problem with all the other fully agentic coding tools is reliability, predictability and speed they function. CWC solves this by requiring more careful workflow - it really pays off the bigger the project becomes.

Key takeaway is that you use long context calls for planning and try to send to the model as little as possible when coding. Even if you have all the bank's money for long context calls you can't buy accuracy and predictability of such calls.

[vivek100]

Agreed, what if we could have a new planning mode, which creates to-do lists with large context and that can be run one by one manually?
Only thing is careful selection of context for subsequent to-dos as they might need code generated in previous to-do.

The idea is can plan diffs in smaller api/chat calls, that is what actually agents do (at least the ones i design do) they pass custom context for each code generation call which kinda increases the accuracy as that individual call triggered via agent . One can even code the custom context builder to reduce the token overload it can just pass the variables and at run time the correct context is passed.

If we go full agent mode its kinda recreating what the IDEs are doing anyway and an overkill, but maybe we could plan better with CWC, instead of one message at a time, we could create plans with tasks for next few hours implement test one by one.

I can try to build this, but want to get your thoughts first.

[robertpiosik]

I think chat providers would not like us doing it... Too much automation for my comfort as my aim is to keep this project alive.

[vivek100]

I understand.