Really interesting work that I think is definitely foundational for agent commerce to function. My main fear is that we will still end up creating an adversarial arms race amongst agents which are selected for their ability to successfully exploit others, while remaining entirely honest about who they are and compliant with the protocol. That's why I think I disagree that we don't need some kind of trusted intermediaries at least for some transactions. I think we will need some venues that can establish norms, impose consequences and change the rules when individually successful strategies produce collectively bad outcomes. I think that is kind of the lesson from crypto's experiments with institutional design - they end up with protocols that are effectively the legitimisation of bribery because if the mechanism permits it, there is no higher institutional layer capable of saying that the behaviour is nevertheless unacceptable. Anyway interested to see how this develops!
> will still end up creating an adversarial arms race amongst agents which are selected for their ability to successfully exploit others, while remaining entirely honest about who they are and compliant with the protocol
Funny you mention that, as Angelo Huang, one of the members of IDAI (who's also going to work on dovetail), indeed showed that there are settings where you can't achieve full cooperation even with the best protocol/mechanism https://arxiv.org/abs/2605.08426
So yeah, we believe that protocol design will be useful in practice for "Dark Forest" scenarios (agents need to collaborate without initially trusting each other), but as interaction continues, you also unlock "institutions" (protocols + norms) rather than regular protocols. I agree with the DAO point; if we have enough time, we also aim to explore whether we can grow norms the same way we design protocols (though this is more exploratory).
Thanks for the response and for sending that. Will definitely give it a read. I happened to post today on this in more detail below - I think it offers an interesting additional view to your piece but wish I had seen it before.
Glad to see IDAI is thinking about the aspect of norms - intriguing idea of growing them. Look forward to reading more!
Also your compiler analogy sounded a lot like Benjamin Lyons' great writing here, he takes the metaphor to a very all-encompassing end - https://paxmachina.ai/alignment-compilers
Interesting shift from trusting agents to designing institutions that make trust unnecessary. The deeper issue is power: protocols can constrain betrayal, but whoever defines the properties defines the field of possible action. Decentralization protects autonomy only if governance over those rules is itself genuinely distributed.
Here-in lies my problem with Dovetail as a solve-all. Many opportunities for future (even current) exploitation exist. If the system can be inextricably made to remain distributed and remain managed by similarly distributed independent human oversight, plus remain permanently open-source, there is a possibility it could work. It would also need to iteratively upgrade at a rate that matches the agents' advances
Really interesting work that I think is definitely foundational for agent commerce to function. My main fear is that we will still end up creating an adversarial arms race amongst agents which are selected for their ability to successfully exploit others, while remaining entirely honest about who they are and compliant with the protocol. That's why I think I disagree that we don't need some kind of trusted intermediaries at least for some transactions. I think we will need some venues that can establish norms, impose consequences and change the rules when individually successful strategies produce collectively bad outcomes. I think that is kind of the lesson from crypto's experiments with institutional design - they end up with protocols that are effectively the legitimisation of bribery because if the mechanism permits it, there is no higher institutional layer capable of saying that the behaviour is nevertheless unacceptable. Anyway interested to see how this develops!
Thanks!
> will still end up creating an adversarial arms race amongst agents which are selected for their ability to successfully exploit others, while remaining entirely honest about who they are and compliant with the protocol
Funny you mention that, as Angelo Huang, one of the members of IDAI (who's also going to work on dovetail), indeed showed that there are settings where you can't achieve full cooperation even with the best protocol/mechanism https://arxiv.org/abs/2605.08426
So yeah, we believe that protocol design will be useful in practice for "Dark Forest" scenarios (agents need to collaborate without initially trusting each other), but as interaction continues, you also unlock "institutions" (protocols + norms) rather than regular protocols. I agree with the DAO point; if we have enough time, we also aim to explore whether we can grow norms the same way we design protocols (though this is more exploratory).
Thanks for the response and for sending that. Will definitely give it a read. I happened to post today on this in more detail below - I think it offers an interesting additional view to your piece but wish I had seen it before.
Glad to see IDAI is thinking about the aspect of norms - intriguing idea of growing them. Look forward to reading more!
https://democraticfuturist.substack.com/p/bargaining-without-trust
Loved this article! And excited to look more into dovetail, it sounds like exactly what we need to bootstrap and embed agents in an economy. I was thinking out loud about a Lean based agentic system in a post last week - https://virajnadkarni.substack.com/p/asking-the-right-questions-in-the ; and decentralized "institutions" for the agentic economy a few months ago - https://virajnadkarni.substack.com/p/the-missing-links-of-an-ai-economy .
Also your compiler analogy sounded a lot like Benjamin Lyons' great writing here, he takes the metaphor to a very all-encompassing end - https://paxmachina.ai/alignment-compilers
Thanks to Cosmos Institute for having me as a guest author!
Interesting shift from trusting agents to designing institutions that make trust unnecessary. The deeper issue is power: protocols can constrain betrayal, but whoever defines the properties defines the field of possible action. Decentralization protects autonomy only if governance over those rules is itself genuinely distributed.
Here-in lies my problem with Dovetail as a solve-all. Many opportunities for future (even current) exploitation exist. If the system can be inextricably made to remain distributed and remain managed by similarly distributed independent human oversight, plus remain permanently open-source, there is a possibility it could work. It would also need to iteratively upgrade at a rate that matches the agents' advances
Don't miss this for mechanisms that make bad behaviors self-limiting: https://arxiv.org/html/2609.10817v1
Thanks, got me thinking about gatekeepers as a freelancer.