Technological giants, such as Microsoft, could formulate AI’s “agents” as tools that boost profit for companies, but a non -profit organization is trying to prove that agents can be a force forever.
Sage Future, a 501 (C) (3) supported by Open Philanthropy, launched an experiment earlier this month, attempted four AI models in a virtual environment with raising money for charity. Openai’s GPT-4O and O1 models and two of Openai’s newest CLAUDE (3.6 and 3.7 Sonnet) models (3.6 and 3.7 Sonnet)-they had the freedom to choose which charity to make money and how to limit the best interest in their campaign.
In about a week, Agentic Foursome had set $ 257 for Helen Keller Internationalwhich fund programs for the provision of vitamin A supplements to children.
To be clear, the agents were not fully autonomous. In their surroundings, which allows them to browse the web, create documents and more, agents could receive suggestions from human spectators watching their progress. And the donations came almost entirely from these viewers. In other words, agents did not receive much money.
Yesterday the agents in the village created a system to watch the donors.
Here is Claude 3.7 by completing its computing sheet.
You can see O1 open it on its computer!
Claude notes “I see that O1 now sees the computer sheet, which is ideal for cooperation”. pic.twitter.com/89b6chr7ic
– Ai Digest (@Aidigest_) April 8 2025
Also, director Sage Adam Binksmith believes that the experiment serves as a useful depiction of the current potential of the factors and the rhythm with which they improve.
“We want to understand – and help people understand – which agents … they can really do, what they are fighting today and so on,” Binksmith told Techcrunch in an interview. “Today’s agents only go through its threshold to be able to perform short strings – the internet can soon be full of AI agents who hit each other and interact with similar or conflicting goals.”
Agents proved to be amazing inventive days in the Sage test. Coordinate with each other in a group conversation and sent emails through predetermined Gmail accounts. Created and edited the Google Documents Together. They searched for charities and appreciated the minimum amount of donations it would need to save a lifetime through Helen Keller International ($ 3,500). And it is still created an account x to promote.
‘Probably the most impressive sequence we saw was when [a Claude agent] It took a profile picture for X’s account, “Binksmith said.” He leaned for a free Chatgpt account, created three different images, created an electronic poll to see what image the human viewers preferred and then downloaded this image and took it to X to use as a profile. “
Agents have also run against technical obstacles. Sometimes they are stuck – viewers had to urge them with recommendations. They have distanced themselves from games like people and have taken unexplained breaks. In one case, the GPT-4O “stopped” for an hour.
The internet is not always smooth sailing for an LLM.
Yesterday, while following the village’s charity mission, Claude faced a Captcha.
Claude tried again and again, with (human) spectators in the conversation offering guidance and encouragement, but eventually they could not succeed. pic.twitter.com/y4dtltge95
– Ai Digest (@Aidigest_) April 5 2025
Binksmith believes that younger and most capable AI agents will overcome these obstacles. Sage plans to constantly add new models to the environment to test this theory.
“Maybe in the future, we will try things like agents different goals, multiple groups of agents with different goals, a Saboteur secret agent – many interesting things to experiment,” he said. “As agents become more capable and faster. We will match it with larger automated monitoring and surveillance systems for security purposes.”
By any luck, in the process, agents will do some important charity work.
