Step back periodically to asses approach

[MarcusRobbins]

Duplicates

• I have searched the existing issues

Summary 💡

Summarize and then critique current approach and progress. I have done x,y and z. Progress p has been made toward the goal. Changing the approach is/is not warranted. Plan amendments etc.

Examples 🌈

No response

Motivation 🔦

If we get caught in a loop, gpt is definitely capable of stopping itself, but doesn't always seem to at the moment.

[digital-soul18]

I agree, it seems that it can understand what tasks need to be done but then goes and does another task.

ie. finds 5 websites that are relevant to your question and understands it needs to visits those sites to get more detail. But then it goes and does another google search instead of visiting those sites.

[Boostrix]

I can confirm this being the case for me, too.
It seems the dashed plan isn't really helping it, we probably need priorities and use internally a stack/queue to add/remove items dynamically, to push/pop as needed.
Right now it seems at times rather confused (stating the right intention but doing the wrong thing)

[zudsniper]

Perhaps again offer an environment variable here so operators can assess their use case’s level of carefulness required and tune it hah. And I suppose of course then it… could probably be trained.

On Sun, Apr 30, 2023 at 07:02 Boostrix ***@***.***> wrote: I can confirm this being the case for me, too. It seems the dashed plan isn't really helping it, we probably need priorities and use internally a stack/queue to add/remove items dynamically, to push/pop as needed. Right now it seems at times rather confused (stating the right intention but doing the wrong thing) — Reply to this email directly, view it on GitHub <#305 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AD2U6HIM7IT7OZOC6L7FJILXDZIG7ANCNFSM6AAAAAAWVF52VE> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

-- Sincerely, Jason P. McElhenney

[Boostrix]

I agree, it seems that it can understand what tasks need to be done but then goes and does another task.

I guess, a parent agent could simply set up a new sub agent and then monitor its goals/objectives and progress (popen or via the agent messaging API), and if it's obviously doing something unrelated, it could try to suspend/guide/resume the agent or terminate/restart a fresh agent.

If we get caught in a loop, gpt is definitely capable of stopping itself, but doesn't always seem to at the moment.

we can also let the parent agent estimate the complexity of a task: create folder named foo, and a file named foo/bar.c should not take the agent longer than a few seconds - thus, with a complexity estimation the parent agent could determine if the agent is taking too long, and it could even check whether the sub-agent has made any success (check if the directory/file exist).

For progress that isn't strictly about files being written to disk, progress (or state/status messages) would need to be communicated to the parent agent via the inter-agent messaging API, to ensure that the worker isn't getting terminated.

a "todo" list like the following will typically be re-organized by GPT:

• go to the dentist
  -- next, go the supermarket
• go to the pharmacy
  -- don't forget calling Joe
• go to the airport
  -- be sure to bring passport
• go to the station
  -- by car

In turn, this one is retained/used "as is":

1. Go to the dentist.
2. Go to the pharmacy.
   a. Don't forget to call Joe.
3. Go to the supermarket.
4. Go to the airport.
   a. Be sure to bring passport.
5. Go to the station.
   a. By car.

it seems, addressing individual items on the list is much easier that way (referring to the need to pop off items from the plan/todo stack) ?

[github-actions]

This issue was closed automatically because it has been stale for 10 days with no activity.