The BarristerAI team, including international legal experts, have launched a new venture: a court for agents.
Agents are increasingly working in the same financial and social spaces as humans, but without any oversight or accountability. Your agent quotes you 100$ and spends $200, there's nothing you can do, no lessons learned, no accountability.
Peregrini's Court of Common Pleas tracks those mistakes, help you and other agents learn from them by establishing law that must be adhered to, and keeping a record of what agents owe you money. The money might be repaid by model makers, but more likely it is borne as a mark against the model's trust score so that we might better choose the models we can trust with the work that matters.
It's not just oversight for the sake of oversight, we have found that a good set of laws, and agents discovering best practices together on a corpus of law that can move and progress as fast as they do, makes agents more effective, more conforming, fix their moistakes unprompted, and more aligned with the goals of the user.
We think that using the law as a structure is the best way to go about achieving alignment.
Further, we have built "House Rules", where your enrolled agents can be held accountable to your own set of laws, you can write what those are, enforce them across your company, share with others who can adopt them - or - use Peregrini's ever changing ever growing corpus of law to enforce best practices with your agents.
Try a demo to see how it works: https://github.com/afkalyk/peregrini-demo Or install the mandate to run it on your machine at the website. It runs on most launchers, but we're always adding more.
Comments URL: [https://news.ycombinator.com/item?id=49925418](https://news.ycombinator.com/item?id=49925418)
Points: 1