Multi-Agent Model

[RobertBateman]

I was reading about the recent research being done around multiple GPT4 agents interacting (article here: https://www.theregister.com/2023/04/11/sims_ai_generation/) and it got me thinking.

Auto-GPT already has the capacity to create sub-agents for tasks, but what if it wasn't a sub-agent? What if it was a co-agent?

Imagine 5 separate autonomous instances of Auto-GPT. What if they were able to collaborate? What if they could share a workspace for files? Or have a chatroom where they could communicate together? It would be essentially the same as 5 humans remote working together.

Now, in this model, each agent could have a role or 'job' which is their main focus, but could be given goals or individual tasks by a supervisor or executive agent.

Basically, TLDR; we should try to re-create a company structure with an executive GPT4 'agent' who the user gives their business goals to, and the executive agent then goes on to 'hire' more GPT4 agents that have autonomy from it, but have their goals set by the executive and report back to it.

The 'executive' may also arrange meetings (chatrooms) where two, or more, Auto-GPT agents discuss their findings/progress.

In particular, I like the idea of having a dedicated 'QA' agent who can make and run test plans (assuming that our goal is the automation of software development here).

My experience of Auto-GPT so far has been hampered by the agents' inability to realize when it has got stuck or notice that it's gone off task. Having multiple agents which could prompt/nudge/question each other could help maintain alignment if one agent goes off on a tangent.

That's enough of my ramblings for now 😅 Thoughts?

[RobertBateman]

Further #showerthoughts on this:

All we actually need is that ability for an agent to interact with a messaging app in order to have multiple autonomous agents.
If we have commands like "start_new_chat" "message_chat" and "read_new_messages" which take parameters for participants and the messages then they have a way to collaborate.
Would probably also need to build in at the base level a way to notify the AI when it has received new messages though. Wouldn't want it to need to actively use a command like "check_for_new_messages". That'd be too slow. You'd want to make a new parameter or field in its self prompting like MESSAGES:"No New Messages", or MESAGES:"1 new message in "ExecGPT-QAGPT-Chat".

Just thinking out loud for now. Might do some coding and see if I can manage it at the weekend or something.

[GoMightyAlgorythmGo]

Further #showerthoughts on this:

All we actually need is that ability for an agent to interact with a messaging app in order to have multiple autonomous agents. If we have commands like "start_new_chat" "message_chat" and "read_new_messages" which take parameters for participants and the messages then they have a way to collaborate. Would probably also need to build in at the base level a way to notify the AI when it has received new messages though. Wouldn't want it to need to actively use a command like "check_for_new_messages". That'd be too slow. You'd want to make a new parameter or field in its self prompting like MESSAGES:"No New Messages", or MESAGES:"1 new message in "ExecGPT-QAGPT-Chat".

Just thinking out loud for now. Might do some coding and see if I can manage it at the weekend or something.

Or if a "meething is not possible" the "save and submit" lets you go back to a point in time and rewrite a promt meaning if the agents are just like chat intances then if you can somehow copy them you can go down a path and when there is a problem you can go back and rewrite the original promt having the information in textform what issues /outcomes happen down the line. I agree that cooperating chats/agents are way stronger. On chatGPT i used to make a 1.0 high lvl plan then let another one take the invidiual steps to 1.1 lover level plan then take those to 1.1.1 chat and 1.1.1.1 and then to a other chats for execution, organisation, and programming chats i also have research chats. I also made meta chats that would tellme what chat to talk to and what promt to use. Automated and with clear structure and careful itterativethinking it is super powerful. But AutoGPT might already be doing it. Someone tht knows all or most of the code or let it run throug chatGPT 3.5t or chatGPT4 and let them tell what it does could make us some pictures, a video or describing how AutoGPT works, that way our theroycrafting and other efforts are better focused and not less effective because having to speculate/ extrapolate/guess.

[SteveHatherley]

I had a similar thought. To have a sort of overseer agent or master orchestration agent that could select from some a few starting prompts when it spun up an agent and then refine the prompt for a specific task. Except I had the thought of using Google Bard as both an research/task/prompt generation/refinement agent and also as a sanity check / orchestration agent for the entire AutoGPT swarm.

I had a pretty interesting conversation with Bard discussing ensemble learning and if it "thought" it could act as an orchestration agent for AutoGPT. Its seemed to think so, and it spat out some code. Unfortunately, I did this from my phone like a dumb@$$ and the chat conversation cleared. Will try to reproduce it and put it somewhere.

On the GPT end of things, I had Bard write an introduction prompt for itself to give GPT, and the GPT also "agreed" that adding Bard agents as both task workers or orchestration or sanity check nodes could be useful. Interesting, the criticism from GPT about Bard was that it was still under active development and might produce inaccurate results. lol

I have a local branch where I've started adding stuff to support a Bard agent as a task worker to start.

[javableu]

I tried to run some conversation like this in chatgpt4, you are a project leader of a group of AI that wants to develop a video game, ...
It looked promising
It could work as long as the coworker builds structured well defined documents that are always the same. If that's not the case, the AI project leader is a bit lost. The main AI project leader could read the report and redistribute the tasks to the team. A kind of giant diagram and allocation of resources might need to be created by the ai project leader.

[GoZippy]

Hey I watched someone using two autoSD agents to talk to eachother to create prompt for SD and then execute the command to generate it ... I was hoping they would share their python hub software they used to coordinate the two and allow independent collaboration of the two agents... have not seen that code anywhere yet... check out this demo https://www.youtube.com/watch?v=EvPHa-CYGXs

[jlafran]

Did you get the code by any chance?

[javableu]

Maybe instead of creating random agent roles, we could do something along the line of :

1 Convert goals and roles defined by the user into a single problem
2 Decompose the problem in x simple parts using fractal tree decomposition of the problem. (you have one problem that you split into two problems and you try to keep the difficulty of the branches constant (Main branch difficulty = child branch task difficulty 1 + child branch task difficulty 2). This should split the problem into simpler tasks.
3 Start from the lowest branches of the trees and ask the ai to solve the task.
4 Update the tree based on the success or not of the agents succeeding at their tasks.
5 If you didn't solve the main problem, repeat phase 3

[javableu]

Given a problem, break it down into two main categories, focusing on tasks that can be executed in parallel. For each category, divide it into two subcategories. Continue this process until the subcategories become simple tasks. Identify any remaining interdependencies among tasks and include these dependencies within the tree structure itself. Finally, create an execution order that takes these dependencies into account, ensuring that tasks with dependencies are performed after the tasks they rely on.

[GoZippy]

Did you get the code by any chance?

Nope, wish he'd share the logic to see...

[GoZippy]

What I really want is to have agent memory persistent options so if her stuck in a loop or main process function freeze we can restart autogpt and it pick right up again.. right now it just restart with the same prompt and redo same work tasks, usually with slightly different results though so nothing ever finishes

[Deno1976]

Almost sounds like building a brain out of Agents

[lee-jian-hui]

so is the multi agent model feasible when using the AutoGPT framework?

[ntindle]

Its work in progress #6896