very interesting read: https://www.anthropic.com/research/multiagent-systems some thoughts: 1. agents need precise, hermetic execution environments when working to avoid pathological trajectories (look up object capability model) 2. agents need better theory of mind — i’ll let the labs do this, my main levers right now would be context engineering (dubious?), or perhaps finetuning (need more gpus tho) 3. proliferation of open-weights models will address the “mode collapse” problem (keep training plz) 4. agents need assets at stake so there are consequences for misbehavior — ethereum is a natural home for this, and already a prototype system can be built (just bond some eth and see what it requires for models to believe there are credible slashing mechanisms, “morality” may need to be baked into the weights, but can start with a strong system prompt) 5. my intuition is that the previous point will quickly benefit from markets/auctions on chain, either for brokering capabilities or as primitives for collective action (“put the swarm in a futarchy” type thing)