Never End

[Roforum]

The discussion between the agents is super funny, it amuses me a lot, especially when it goes into a loop. Then you give them a task and they start discussing something else.
I'm using the ollama version and so far I haven't been able to finish anything, I left it to a small task to see when it finishes, 26 hours have passed and it's still going.

I think that in many aspects it is much better than Davin or CrewAI, but it needs to be tickled a little to complete a task.

Then the option for long term memory should be added, which saves everything in the /data folder in the same folder where the python file runs and for each task to create a new subfolder in the /data folder.

[Left4haxx]

Same, one time the agents started speaking spanish. Also used ollama but with phi3.
I think this could be addressed if there was a moderator agent. Messages could be logged and store information such as :

• Current Task Objective
• Message
• Time Sent
  Or something of the sort. The purpose is to have a trigger that either halts the program, or intilaize the moderator agent in parallel.
  The moderator agent could have a module file(following Ollama theme) prompting it to act as a moderator that checks if the conversation is related to the primary objective. This issue and solution needs to be thought out more, thus I am just giving an opinion on how we can approach this.

[mfp20]

Same here. I'm asking "create a snake game using HTML, CSS, JS." and 4 times out of 5 they keep talking and cheering each other on how good the design rules are, slowly shifting on more general terms until they start speaking about spacetemporal analysis and ... if I let them go ... I might get the blueprints for a time machine, or teleport, maybe, in the next days ... but I was bored and cut them off after 2 days only. Tried different models (llama3, codelama, qwen2, codewen, deepseek-coder, deepseek-coder-v2, dolphincoder, ...), and tried different prompts, including modifing the ones supplied in the python scripts. No improvements.

The other 1/5 attempts, I got empty files and folders or the files and code output in console in json format; partial success has been adding the UTF8 directive on file opens.

The whole thing looks like the callcenter of a big company: dumb operators wasting your time with politeness and useless procedures until you get a useless advice :)

That said, please keep trying because the thing is bloody sexy!