What is the limit of long Agent tasks?
What I gained after running the codex goal for more than ten hours.
I have also thought about this problem before.
Is a simple agent loop plus a very large step run number a long task?
What kind of task is worth a long task?
How long is a long task?
A long task consumes so many tokens and so many rounds.
In the second half, its memory, context compression, and fitting degree are all big problems?
How to solve it?
After Codex goal was launched, I started to experience it as soon as possible, and also shared several prompt word templates.
Let me talk about some of my feelings: A good long task must have strong model capabilities, such as millions of contexts, super coding capabilities, reasoning capabilities, and planning capabilities.
Otherwise, it will really be a waste of tokens.
Don’t try it easily if there is no need.
Personally speaking, long tasks are suitable for pioneering tasks.
You don’t have enough knowledge and experience to constantly tell and correct the behavior of the big model.
The best way is to let it explore and you do a shallow acceptance.
For example, after running for a few hours, it has made many attempts on that path according to the path and strategy you set at the beginning.
At this time, the big model will tell you very firmly that that method is limited.
For example, I started running before going to bed.
After I woke up, a codex goal of more than ten hours sent me a long paragraph of its various explorations.
Finally, it told me that this method had reached its limit, and the benefits of further methods were very small.
I have never experienced this experience before running such a long task.
It is not simply a type of failure, it is a conclusion after the methodology is fully verified, with process and output.
I think this is one of the values of long tasks.
Of course, following the above question, an external experience library is needed to take over the products of these processes, because as the next goal starts, it will undoubtedly lose a lot of valuable experience and lessons.
One of the best practices I have discovered is that every long task is worth your extra time to proactively accumulate some high-value content.
Another suggestion is to have some good skills to assist in calling tools.
This is where the most time and tokens are consumed.
Overall, not all tasks are worth using long tasks.
Most of my goals can be completed within half an hour.
Because the scope and acceptance criteria are limited, they must be written clearly.
You can read what I wrote in my previous post.