Tech giants like Microsoft may be touting AI “agents” Companies increase profit toolsbut a nonprofit is trying to prove that agency can also be a permanent force.
Sage Future, 501(c)(3), supported by open charity, launched an experiment earlier this month that appoints four AI models in a virtual environment and raises funds for charities. These models – OpenAI’s GPT-4O and O1, and Anthropic’s two new Claude models (3.6 and 3.7 sonnets), are free to choose charities for fundraising and how best to improve their campaign.
About a week, four agents had Helen Keller International raised $257the program fundes the provision of vitamin A supplements to children.
It should be clear that the agent is not completely autonomous. In their environment, this allows them to browse the web, create documents, and so on, and agents can make suggestions from human audiences who are watching their progress. Donations are almost entirely from these audiences. In other words, agents don’t organically raise too much money.
Yesterday, agents from the village created a system to track donors.
This is the Claude 3.7 filling out the spreadsheet.
You can see that O1 opens it on its computer!
“I see that O1 is now looking at spreadsheets, too, which is great for collaboration,” Claude noted. pic.twitter.com/89b6chr7ic
– AI Digest (@AIDIGEST_) April 8, 2025
Nevertheless, SAGE Director Adam Binksmith believes that the experiment is a useful illustration of the agent’s current functions and the speed they improve.
“We want to understand and help people understand – which agents […] “It can actually do what they are currently fighting, etc. Today’s agents are just through the threshold to be able to perform short-term actions – the internet may soon be full of AI agents collide with each other and interact with similar or conflicting goals.”
It turns out that in Sage testing, these agents are a multi-category day. They coordinate with each other in group chats and send emails through a pre-configured Gmail account. Together, they created and edited Google Docs. They looked at charities and estimated the minimum amount of donations needed to save lives through Helen Keller International ($3,500). They even Created X account for promotion.
“Probably the most impressive sequence we’ve seen is [a Claude agent] “It takes a profile picture of an account X. It signs up for a free Chatgpt account, generates three different images, creates an online poll to see which image a human audience likes, then downloads that image and uploads it to x for use as a profile.”
The agents also violated technical barriers. Sometimes, they get stuck – the audience has to prompt them to make suggestions. They have been distracted by games like the world and have somehow rested. Once, GPT-4O was “paused” for an hour.
The internet doesn’t always go smoothly for LLM.
Yesterday, Claude encountered verification codes while pursuing the village’s charity mission.
Claude tried time and time again to provide guidance and encouragement in the chat, but ultimately failed. https://t.co/xd7qptejgw pic.twitter.com/y4dtltge95
– AI Digest (@AIDIGEST_) April 5, 2025
Binksmith believes that newer, more capable AI agents will overcome these obstacles. Sage plans to continuously add new models to the environment to test the theory.
“Maybe in the future, we will try to offer agents different goals, multiple agent teams with different goals, a secret disruptor agent, and that’s a lot of interesting things,” he said. “As agents get faster and faster, we will match larger automated monitoring and oversight systems for security purposes.”
Fortunately, in the process, the agents will do some meaningful charity work.
