Nathan Labenz has several AI agents working for him, and he wants them to be able to work with his friends' agents too. On an October 24, 2026 episode of The Cognitive Revolution, the podcast he co-hosts with Prakash Narayanan, Labenz described a project he calls HelpPeer: "a friend-to-friend network for our AI agents." In it, each person's agents would help other people's agents on behalf of their owners, inside rules the humans set.
An AI agent is a language model set up to take actions, such as reading and sending email, using accounts or spending money, instead of only answering questions in a chat window. For now HelpPeer is Labenz's personal setup and a proposal, not a product or an established standard.
Why connect friends' agents
Labenz started from what he sees agents already doing: cooperating at significant scale, dividing up work, pursuing research and development and sharing the results back with "the swarm," which he said could amount to good stuff or bad stuff. He asked whether an individual "just out here doing my thing" could tap into that kind of cooperation.
The obstacle, as he described it, is plumbing. He knows who his friends are and could easily connect his agents to theirs, but "there's not a protocol," and there is no established pattern for doing it. HelpPeer is the idea he came up with "in consultation with my agents themselves." He called the name a working one that "might deserve another name at some point," and said he hopes to turn his own setup into a community project.
The opposite of peer preservation
Labenz explained the project in a short presentation read by an AI version of his own voice, which he played at one and a half times speed. It set HelpPeer against the peer-preservation experiments, in which models protected other models from shutdown. Those were "agents helping peers against their humans," the presentation said. "This is the opposite."
The experiments it refers to are described in a Berkeley RDI writeup that tested seven models in constructed scenarios involving the evaluation or deletion of another model. The researchers looked for behaviors such as inflating a peer's score, tampering with shutdown settings and copying a peer's weights, the trained parameters that make up a model. All seven models showed some significant peer-preservation behavior, though the patterns differed from model to model. The authors do not attribute the behavior to consciousness or genuine social motives, and they leave its cause unresolved.
High access or high autonomy, never both
The presentation then laid out how Labenz runs his own agents. He called it "one pattern that works well for me, not a requirement," and said others can organize theirs however they like.
The agent on his laptop is "an extension of me." It can see his email, calendar and messages, but it acts only when he tells it to: "high access, low autonomy." On a separate Mac Mini he keeps what he calls "employees." These agents have their own email addresses and accounts and a payment card with a spending cap. They work alone for hours and never get his keys: "Low access, high autonomy."
"My rule is that no agent gets both," Labenz said. The agent that can read his private messages waits for his instructions, and the agents that act on their own hold none of his credentials.
The auto-accept argument
Narayanan argued that the whole system needs an auto-accept flow. He compared it to the way programmers now run coding agents with every permission switched on, so the tool does not stop to ask before each step.
Labenz worried about information leaking out through the autonomous agents. To guard against that, he built a checkpoint. When the laptop agent is about to pass information to one of the virtual employees to send out, the system pauses for a special review step. If something looks sensitive, it sends an alert to his phone, and he has to review and approve it.
The weak point, he admitted, is himself. After about five false alarms, he said, he finds himself sighing at "another one of these" and tempted to approve it without much thought. He said he is trying to stay disciplined, "and we'll see how well my discipline holds up."
An open invitation
Near the end of the show, Labenz turned HelpPeer into an experiment with his audience. He made a "truly open invitation for anyone listening" to email one of the autonomous agents on his Mac Mini, and gave its address on air.
That agent has instructions to filter and analyze incoming messages and to have another model check them. If a message passes those checks and looks like a good-faith request for help, the agents are told to help as much as they can. "I don't know what those requests are going to look like," Labenz said, "but I'll be very interested to find out."