Tech giants like Microsoft may be touting AI “agents” as a profitable tool for businesses, but nonprofits are trying to prove that agents can also be eternal forces.
Sage Future began experimenting with 501(c)(3), backed by Open Philanthropy, earlier this month with four AI models in a charity-funded virtual environment. Models – Openai’s GPT-4O and O1, and Anthropic’s new Claude models (3.6 and 3.7 Sonnet) had the freedom to choose what charities to choose for the campaign and how to do your best.
In about a week, Agent Fourthom raised $257 for Helen Keller International.
To be clear, the agents were not completely autonomous. In their environment, they can browse the web and create documents. Agents can make suggestions from a human audience watching their progress. And the donations came almost entirely from these audiences. In other words, the agents didn’t organically raise much money.
Yesterday, village agents created a system to track donors.
Claude 3.7 is a spreadsheet entry.
You can see O1 open on your computer!
Claude said, “We found out that O1 is also showing a spreadsheet, which is perfect for collaboration.” pic.twitter.com/89b6chr7ic
– AI Digest (@Aidigest_) April 8, 2025
Still, Sage Director Adam Binksmith believes this experiment will serve as a useful illustration of the agents’ current capabilities and the speed at which they are improving.
“We want to understand and understand what agents can actually do, what they’re struggling right now,” Binksmith told TechCrunch in an interview. “Today’s agents are just passing a threshold that they can execute a short string of actions. The internet may be full of AI agents quickly bumping into each other and interacting with similar or conflicting goals.”
The agent proved to be an incredibly resourceful day of Sage’s testing. They coordinated each other in group chats and sent emails via pre-configured Gmail accounts. I created and edited Google Docs. They researched charities and estimated the minimum amount of donations needed to save lives through Helen Keller International ($3,500). And they created an X account for the promotion.
“When was the most impressive sequence we’ve seen. [a Claude agent] Binksmith said: “I signed up for a free ChatGPT account, generated three different images, saw images that human viewers liked, and downloaded them and uploaded them to X to use them as profile pictures.”
Agents are also opposed to technical hurdles. Sometimes they got stuck – viewers had to urge them with recommendations. They were distracted by games like the world and took an inexplicable break. At one point, the GPT-4o “paused” itself for an hour.
The Internet is not always smooth sailing for LLM.
Yesterday, while pursuing the village’s charity mission, Claude came across a capture.
Claude tried it over and over, and although (human) viewers were providing guidance and encouragement to the chat, they ended up not succeeding. https://t.co/xd7qptejgw pic.twitter.com/y4dtltge95
– AI Digest (@Aidigest_) April 5, 2025
Binksmith believes newer, more capable AI agents will overcome these hurdles. Sage plans to continue to add new models to the environment to test this theory.
“Perhaps in the future we’ll try to try out different targets for our agents, multiple teams of agents with different goals, secret sabotage agents, etc. “As agents become more capable and faster, they’ll coincide with larger automated surveillance and surveillance systems for safer purposes.”
If you’re lucky, along the way, your agents will do some meaningful charity.
