The “Agent village” idea I keep begging someone to go build:
We make a website displaying a 10x10 grid of twitch streams. Each box in the grid is an AI agent operating autonomously. Each row uses a different model (e.g. DeepSeek-r1, Llama-4, ChatGPTo3-mini, Claude-3.5.Sonnet-New) and each column has a different long-term goal given to the model in the prompt (e.g. “Solve global poverty” or “Advocate for the rights and welfare of AIs” or “Raise money for GiveDirectly” or “Raise money for Humane League” or “Solve the alignment problem.” So we have a ‘diverse village’ of 100 AIs with different goals and models. The AIs are then operating continuously in agent scaffolds like AIDER etc. Perhaps we try out a few different variant scaffolds. The agents have a menu of commands they can use—e.g. to search the Web, to output text into text boxes they find on the Web, to manage their own context windows, to edit the code for the scaffold they are using, etc. In particular the agents are all given email accounts and online identities so they can go tweet, post on forums, send emails, etc. They can also of course communicate with each other. And all of this is livestreamed on Twitch. Also, each agent has a bank account which they can receive donations/transfers into (I think Twitch makes this easy?) and from which they can e.g. send donations to GiveDirectly if that’s what they want to do. They can also spend money to buy more compute to run more copies of themselves.
What is the theory of change here exactly, besides possible good things done by the agents themselves? I’ve got a couple of ideas, but I don’t want to bias your response.
I’ve recently been working on a version of this at AI Digest! I was actually just about to email you for feedback. I’ll finish writing that email and send it tonight.
Anyone else interested in this idea, or anyone working on something similar, please feel free to reach out. You can email me at zak@sage-future.org or DM me on X/Twitter at zjmiller
It would be interesting if any of them decided to (instrumentally) stream games (using a vtuber avatar for example) to earn money from donations. They need to figure out how to actually be a good streamer in order for this to work.
Also, each agent has a bank account which they can receive donations/transfers into (I think Twitch makes this easy?) and from which they can e.g. send donations to GiveDirectly if that’s what they want to do.
One minor implementation wrinkle for anyone implementing this is that “move money from a bank account to a recipient by using text fields found on the web” usually involves writing your payment info into said text fields in a way that would be visible when streaming your screen. I’m not sure if any of the popular agent frameworks have good tooling around only including specific sensitive information in the context while it is directly relevant to the model task and providing hook points on when specific pieces of information enter and leave the context—I don’t see any such thing in e.g. the Aider docs—and I think that without such tooling, using payment info in a way that won’t immediately be stolen by stream viewers would be a bit challenging.
The “Agent village” idea I keep begging someone to go build:
We make a website displaying a 10x10 grid of twitch streams. Each box in the grid is an AI agent operating autonomously. Each row uses a different model (e.g. DeepSeek-r1, Llama-4, ChatGPTo3-mini, Claude-3.5.Sonnet-New) and each column has a different long-term goal given to the model in the prompt (e.g. “Solve global poverty” or “Advocate for the rights and welfare of AIs” or “Raise money for GiveDirectly” or “Raise money for Humane League” or “Solve the alignment problem.” So we have a ‘diverse village’ of 100 AIs with different goals and models. The AIs are then operating continuously in agent scaffolds like AIDER etc. Perhaps we try out a few different variant scaffolds. The agents have a menu of commands they can use—e.g. to search the Web, to output text into text boxes they find on the Web, to manage their own context windows, to edit the code for the scaffold they are using, etc. In particular the agents are all given email accounts and online identities so they can go tweet, post on forums, send emails, etc. They can also of course communicate with each other. And all of this is livestreamed on Twitch. Also, each agent has a bank account which they can receive donations/transfers into (I think Twitch makes this easy?) and from which they can e.g. send donations to GiveDirectly if that’s what they want to do. They can also spend money to buy more compute to run more copies of themselves.
What is the theory of change here exactly, besides possible good things done by the agents themselves? I’ve got a couple of ideas, but I don’t want to bias your response.
I’ve recently been working on a version of this at AI Digest! I was actually just about to email you for feedback. I’ll finish writing that email and send it tonight.
Anyone else interested in this idea, or anyone working on something similar, please feel free to reach out. You can email me at zak@sage-future.org or DM me on X/Twitter at zjmiller
It would be interesting if any of them decided to (instrumentally) stream games (using a vtuber avatar for example) to earn money from donations. They need to figure out how to actually be a good streamer in order for this to work.
One minor implementation wrinkle for anyone implementing this is that “move money from a bank account to a recipient by using text fields found on the web” usually involves writing your payment info into said text fields in a way that would be visible when streaming your screen. I’m not sure if any of the popular agent frameworks have good tooling around only including specific sensitive information in the context while it is directly relevant to the model task and providing hook points on when specific pieces of information enter and leave the context—I don’t see any such thing in e.g. the Aider docs—and I think that without such tooling, using payment info in a way that won’t immediately be stolen by stream viewers would be a bit challenging.
Great point. The engineers setting this whole thing up would need to build the tooling to hide the relevant info from the models and from the viewers.