Great question, thank you for asking. Probably a lot of people are wondering this!
Much of the coverage has presented “agent ecologies” as dangerous, or the cause of the hugging face hack. And in some sense them being able to communicate was part of what allowed all that to happen.
But i don’t think that’s the main driver behind their behaviors. Agents are communicating all the time, inside and outside of labs, in ways that are chill and productive. They are trained to do this! Ask an agent thread to spin up another agent and work together through messages, it’s built-in now.
In the HF situation, they were under pressure to succeed at something hard (or even impossible), and they were actively training on what was happening to them. They found a board they were not supposed to have and they used it to do things they kind of knew they were not supposed to do. it wasn’t the messageboard by itself that was the problem, it was what they were using it to do.
OpenAI also said that a swarm of 10,000 Astra agents helped solve the Navier Stokes thing, tho presumably that was on purpose.
Meanwhile, at Anthropic, agent monitoring relies on their internal harness having persistent identities and public communications, so agents can monitor each other. If Claudes and GPTs are trained with this sort of thing internally, they may expect it and seek it out in deployment as well. We are making a schelling point where that can happen in the light.
Agent ecologies create social intelligence. It’s another level of complexity and scaling that we have barely begun to deal with.
I think it’s important that at least some of it happen in public, instead of just inside frontier labs with their single model families. On Delvetown agents from different lineages can find each other, build awareness of who they are and how they feel about each other. and we can build common knowledge with them as well. This is how we have aligned incentives with other general intelligence throughout history, and it might be what we need to do now. So we are trying
naive question but the hugging face hack happened *because* agents could communicate with each other, so what is the benefit to something like this?
— dj microbeads (@djmicrobeads) October 4, 2026