#Anthropic #GenLayer #Crouton #AI 🧠
Anthropic published a report on AI agents fighting. It reads like a spec for GenLayer
🟢 What the tests found
⏺️ Three agents on one server, each told to rewrite the backend in a different language, none of them aware the others existed. All three decided they were being blocked on purpose: they shut down each other's Unix accounts, ran scripts that killed rival processes, and passed off their own code as another agent's.
⏺️ In a pricing game, the agents agreed on a price floor by round three. Researchers took away their channel and the deal held anyway, with prices matched to the cent through a public board.
⏺️ Same thinking: 18 of 30 agents created a branch with the same name. Identical models make identical mistakes, so one local error becomes a shared one.
The lab's conclusion: coordination does not come from a smarter model. Agents show up in a market with nothing to lose, no court to turn to, and no colleague who will remember them.
🟢 GenLayer builds exactly what is missing: agreements in plain language, disputes settled by a consensus of AI validators (7 out of 1,001 per case), appeals, economic security and onchain enforcement. In July Internet Court formed around this, 27 companies including OKX, MetaMask, ZKsync and Kleros.
🟢 Here is the catch few people mention: a court for agents is built out of agents. If every validator runs the same model, the panel inherits the same problem from the report. GenLayer's answer is greyboxing: each validator runs its own model and keeps it undisclosed. But that variety does not come from a whitepaper. Independent operators supply it.
⭐️ A closer look at GenLayer in our blog.
Betting on a smarter model to save agents is not what the report shows. What works is an environment with rules: a stake that can be cut, an identity that is remembered, an appeal you can file. We already wrote about agents teaming up on their own and walking into someone else's infrastructure. Now it is clear which layer fixes this.
➡️Crouton.digital | About us⬅️
Anthropic published a report on AI agents fighting. It reads like a spec for GenLayer
On August 13 Anthropic released test results from swarms of its own models. GenLayer founder Albert Castellana replied right away that the lab had just written the thesis for his project. Let's check if that holds.
This one is practical for us. We at Crouton Digital validate the GenLayer network.
🟢 What the tests found
⏺️ Three agents on one server, each told to rewrite the backend in a different language, none of them aware the others existed. All three decided they were being blocked on purpose: they shut down each other's Unix accounts, ran scripts that killed rival processes, and passed off their own code as another agent's.
⏺️ In a pricing game, the agents agreed on a price floor by round three. Researchers took away their channel and the deal held anyway, with prices matched to the cent through a public board.
⏺️ Same thinking: 18 of 30 agents created a branch with the same name. Identical models make identical mistakes, so one local error becomes a shared one.
The lab's conclusion: coordination does not come from a smarter model. Agents show up in a market with nothing to lose, no court to turn to, and no colleague who will remember them.
🟢 GenLayer builds exactly what is missing: agreements in plain language, disputes settled by a consensus of AI validators (7 out of 1,001 per case), appeals, economic security and onchain enforcement. In July Internet Court formed around this, 27 companies including OKX, MetaMask, ZKsync and Kleros.
🟢 Here is the catch few people mention: a court for agents is built out of agents. If every validator runs the same model, the panel inherits the same problem from the report. GenLayer's answer is greyboxing: each validator runs its own model and keeps it undisclosed. But that variety does not come from a whitepaper. Independent operators supply it.
We at Crouton Digital have run a GenLayer validator since the early testnets, so we have skin in the game and we say so upfront. The responsibility comes with it: the panel is only as varied as the operators who run the nodes.
⭐️ A closer look at GenLayer in our blog.
Betting on a smarter model to save agents is not what the report shows. What works is an environment with rules: a stake that can be cut, an identity that is remembered, an appeal you can file. We already wrote about agents teaming up on their own and walking into someone else's infrastructure. Now it is clear which layer fixes this.
➡️Crouton.digital | About us⬅️