Legal Infrastructure for Agent Transactions – San Francisco 2025

hi I'm dazer Greenwood and I'm really into generative AI uh my area is law um practice law in the 90s Tech Law and um now I'm back in technology um and I run a research group at MIT called law. mit.edu where we do computational law prototypes and um you know convening and uh I live in Oakland for the last few years and now I'm happy to be affiliated with Stanford law school as well something called codex which is the computational law kind of um counterpart here in the uh bay area and I'm really into agents and this is the reactions I usually get when I talk about AI agents and I think somebody in the couple of panels ago said yeah we should connect climate change and AI because they're both things that are going to kill us um so that's the person on the left um and there's also like hey agents those things when we configure them right and align them could really help us that would be the person over here and um I've actually been into this for a long time I think agents are all the rage right now oh hold on I need to put a timer on her we'll be finishing about 4m. okay um in the late 90s um I was working a lot on uniform law you guys probably heard of the Uniform Commercial Code that's the main commercial law in the United States at the state level um when e-commerce was getting to be a big thing in the late 90s we realized we needed some legal updates and reforms to make sure we could have electronic records electronic contracts that was called uniform electronic transactions act I was deeply involved in drafting that and we actually I was this memo here shows some of the work that I had done making sure we had electronic agents Provisions in that law and we do I'll get into those in a moment when we talk about the legal infrastructure supporting electronic agents uh and llm agents now but the main thing is that um these Provisions were um uncannily preent I think in terms of how autonomous systems have are playing out today and um over the years I've done a lot of stuff with um autonomous organizations and different methods of wrapping agents the the main thing that's different now than what we um really had at hand even just a few years ago yes the question was could I Define what I mean by agent and that brings us exactly to this slide in the next one so perfect timing um in the old world you could say as in you know a few years ago and before um you know we certainly had agents and um but they were more let's call it deterministic systems that could do a bunch of things automated um workflows um but what's different now uh in particular with the new class of Transformer modelbased generative AI is you actually can get a kind of almost like reasoning or a kind of thinking it's not the same way humans do it but based on kind of benchmarks and evals and just lived experience using these systems they can look at novel things that no one had ever seen before and come up with potentially reasonable inferences about that and when you start to you can compose questions and uh prompts and other systems like the ability to use apis and to store memory over sessions and access other data with a lot of application logic um you can actually sort of um compose the these capabilities to be kind of like an agent that can think um this is really interesting especially in a legal context when you can see well first of all practically um you know doing automated Contracting systems every single transaction code and everything else had to line up perfectly it was like an engineered deterministic system for Supply chains payment networks everything like that even for smart contracts in the blockchain context now but most contracts are a little bit more ambiguous you know they're a little bit more fuzzy um and it doesn't take necessarily genius level of like you know um federal judge skills to figure out what the contract means it is actually well within the capabilities of especially the most current llms Claud 3.7 GPT 4.

5 Gemini 2.0 pro Etc uh and some of the open source models as well so what's different is they can kind of like quote think for themselves and they can adapt so this one hopefully will scratch the itch of the question I think some of the capabilities that distinguish these things are you can start with a goal you kind of give it a goal that it's doing like conduct a transaction brainstorm something take a look at this environmental remediation plan and identify if it's in alignment with these 10 priorities or not Etc it can inv it can deal with a lot of environmental comp complexity um like inputs that weren't pre-programmed and that might be changing can adapt and it can start to even execute on things so in a chat interface the execution is that it gives you an put that's like language or or an image but as you start to encapsulate these things in applications the output can absolutely be much more sophisticated here's something from just last year that I I really like from a research perspective from Sakana labs in Japan they put together a bunch of agents in these different Lanes to um generate ideas and and they can generate ideas when you prompt them right and give them access to like archive.org and and other systems it generate ideas for research create synthetic data sets create some you know do some software engineering some Python and some other stuff numpy scipi run experiments based on hypotheses it generated get results another set of Agents look at those results is there anything interesting here you know no no no no yes yes no maybe can start to rate and rank those things which they do um and then it could take the one that's really good and then another set of Agents we're like okay let's look at results let's look at our hypothesis let's here's templates for research papers of this type let's write a draft of a paper let's iterate the paper do stuff like deep research which we have now from open Ai and and others and actually publish a paper um not peer reviewed not even human reviewed um and so this isn't necessarily something that we're going to you know want to do right now but as a proof of concept for how sophisticated these things can be I think it was very good and the code is available to download and to run yourself on GitHub it it's costly there's a lot of API calls but this to me was astonishing and this is like almost a year ago now um within the last few days we're seeing much more sophist over the year we've seen a steady pump of incremental and sometimes bigger increases in what types of AI agents we have over the last few days we had Manis out of China a general purpose agent the first one I showed you is domain specific they're all the prom and everything is connected to doing research and generating a research thing and writing a paper um this one's mana's general purpose so you can throw any goal or idea at it it will reason about what it is and start to um produce results for that idea including generating code or coming up with a business plan or just whatever it is um on the right side um just a few days ago open aai finally dropped their agents uh the open AI agent AI agent SDK I was really fortunate to have early access to that and it is phenomenal um at the bottom there's a uh a link to or a um screenshot from something at dazzer greenwood.com which you can click on if you're interested to learn more but that's basically demo Knight of some of the stuff you can do with this new agent API I did something like Manis but it only took me like not very long to do that so when you when you start to create better tooling and better um developer experiences for these agents they're huge now you can use these things for democracy you can use these things for collaboration a lot of good stuff the thing that I'm particularly interested in um is a little off to the side but it is not without public um and civic interest and need that's transactions it's a kind of lie practi it's kind of what I'm into kind of purchase and sale e-commerce and that is part of almost everything that we do you know even even if you're running a environmental nonprofit at some point you want to cut a deal you want to Source stuff you want to have an arrangement with another Consortium like you know these interactions and transactions and legal stuff is part of it it's kind of a hardcore use case um some of the basic normal ones from UCC are like um you know shopping travel booking and these are these are very well supported in the um current um transactional agent um Frameworks and service um some Mass Market ones that are now available as of a few months ago are um perplexity which has something called buy with perplexity so when you do your search your AI search with perplexity and you're looking for products or services or what have you um as a pro user it'll give you a a set of answers and you could click on it and it'll give you prices you can select a product you want go and and basically it'll purchase it for you and send it to your ship to address um pretty interesting I've I've used it um and it's it was pretty good um stripe is even more mainstream than perplexity and so a couple three months ago they dropped the stripe stripe agent API and so they now have built out this entire part of the stripe API stripe is facilitates a huge amount of commercial payments payment transactions right now and um they made it possible to basically um scope authorization spending limits and other dimensions of a of um author of delegated authorization and basically allocate that those API rights and authorizations to agents okay stripe is Big so like this is not a futuristic talk anymore the way it's been most of my life like this is happening now but you know I don't think we're completely completely ready for this going back to the late on the right going on the roller coaster like I do have some trepidations about this and I feel like there's some gaps that need to be filled and Commons are a very good way to do some of this um they will be filled one way or another they'll be filled by companies um that are optimizing for their interests um but we can get ahead of this a little bit so one of the things that um needs to happen is authenticated delegation for authorized AI agents operating on behalf of a person or your organization this is not impossible but you know there's some dots to be connected to do that um and so there's a great paper if I may say so as a co-author but I didn't come up with this idea I kind of came along later um for authenticated delegation that uses open ID connect and ooth 2 and basically configures these existing widely used standards so that some of the off tokens and so forth are allocated to agents and you could and the user maintains control the whole way way through and the external parties can verify and rely on that there's also a whole bunch of existing laws I mentioned earlier that we can look at and kind of reframe for these new issues um uniform electronic transactions Act is is the biggest unknown law out there that I want to make sure everyone is at least heard of um and so this is enacted in 49 states of the United States for like 20 plus years it's part of the legal fabric and all the territories of the United States it's part of the legal fabric of the United States and it has Provisions for autonomous agents okay it was a little ahead of its time and I remember I showed you that memo like what we were advocating that we make it future proof to include deterministic agents and autonomous agents we did that I assumed it would have happened you know 20 years ago it was always around the corner well it's happening now it's happened now um and they basic Bally come up with some specific um well first of all they Define it as a computer program that's used independently to initiate basically a transaction and it can finish the transaction without review or action by the individual okay so that's like the essence of autonomy no human review or approval and it conducts and finalizes a transaction it's like legally enforceable okay it does automated transactions so we have some other Provisions that enable and and Define when when you have agents on both sides of the transaction which is legit it's good it's inevitable and really interestingly um we also put some Provisions in there for what we foresaw especially with these generative non-deterministic systems that there'll be errors introduced when you can't know exactly what an llm is going to come up with in advance it could come up with things that may stray from your intention and may not be aligned with exactly what it is that your interests are um there's a I think there's a lot of ways to deal with that including having the providers have a fiduciary duty to the user and we can get into that more later but technically in the law right now if an agent system does not provide the ability if an if there's an error that results from the agent doing stuff that got out of alignment with the user thanks um the user can reverse the transaction under this provision uh section 10 B of UA if the agent system did not provide a means to prevent or correct the error it's very fair um and this this was standard back in the late 90s when we wrote this in 2025 this would have been hard to get through because industry would have been oh my God you can't reverse the transaction but you know hey to suck it up it's the law and you know this is the law of the land this is really handy we can use this law um you can read more about that including ways to support and reflect it within systems that use transactional agents at daag greenwood.com and kind of break it all down I also we also have an open source reference implementation for how to do error detection and to provide um error prevention or error correction is sometimes needed because you can't you don't know what it's going to do in advance so that you users so it's fair so the user can basically hit the red button on a transaction before they're on the hook for it and also the merchant or whoever's on the other side doesn't have to worry about this lingering uncertainty about will the user later say it wasn't what I intended and under u i can reverse the transaction no just build it so that you gave them an opportunity to prevent or correct and so that's what this open source does just to show people how to do that it's actually not that hard when you think about it but you do have to think about it and first you will have to have heard of it therefore I'm telling you what does this have to do with the comments everything um so like I do a lot of Consulting work for companies and um the way this is going as I was mentioning is going to be effective and productive and um efficient uh but it may not exactly support and reflect the best interest of individuals and um consumers and so forth and so I've been putting some time a lot of time into working with state legislator to develop um updated legal Frameworks um that fill gaps that that U had a a statute I'm really proud of is in Wyoming it's the digital uh it's the personal digital identity stat statute um and it creates definitions for these new kinds of Agents um more more recently um within the last year I've been helping Wyoming with legislative hearings on llm agents and um most importantly right now I think the best thing to do is to get together as people and to come up with best practices model statute and maybe most practically model contract provisions and configurations for systems that use AI agents that can create um fair and and um beneficial results this is great technology I'm definitely more like the lady on the left when it comes to this I love this stuff it's just what I've the tech I've always wanted especially for legal stuff and transactions you just have to do it in a good way and a good way to do it in a good way is the Republic Commons so therefore I'm convening um with consumer reports and with um hii and Stanford codex at um Stanford um a a new work stream uh where we come up with legal um Frameworks or mostly adapt and apply legal Frameworks they support and reflect the use the imminent emerging and going to be humongous by the end of this year use of AI agents so we'll get out ahead of it we'll sort of have some reference implementation some simple protocols that basically just apply existing protocols that people already use and some ways to structure your contract so everything is hunky dory um I wish I could get into more detail um oh well the most important thing is if you're interested in this or you think you know someone who might be your lawyer people your entrepreneurs your Visionaries um send them to computational law.

org and there they will see some really awesome stuff where they can basically find out about the first big open Workshop that we're doing on April 8th as part of future law week at Stanford and all they have to do oh this is outdated anyway all in the new version um they will have a way that they can sign up uh for oops I'm going to stop clicking buttons they can sign up and request the uh request and invitation of this Workshop um and even if you can't make the workshop although it's in person and virtual they should if they're interested in this stuff just put the form in and say I was at funding the commons or someone at funding the commons told me about this and um I want to get involved like put me on your email list like let's set up a call so I'm in active community building mode right now and um with a little bit of resource we can go a long way so I hope that you're interested in AI agents and would like to build the future that we would prefer to inhabit by coming together in a common way now thanks [Applause] do we have time for a question or am I just so awful oh good I'm not that awful in terms of time anyone have questions or ideas yes sir yeah um so the question was what's the scope of application for this stuff I'm super coming like I you had me at web 2 so I'm like kind of like very much there still um but when I think about a solidity contract or or anything that's a smart contract um which is good and it's important I hope that continues to develop the way I would apply this and actually there's like a branch in the GitHub repo for the open source reference implementation that you'll see that does this is I would lean on error prevention because there ain't much you can do with error correction you can sort of do outof Channel stuff and but it's a lot it all sounds scary and terrifying to me it's lady on The Right Stuff um so I think lady on the left stuff is error prevention and um it's super doable actually the stuff isn't that hard we just have to think it through and do it anybody else yes sir State oh New York those bastards because they're the Empire State what's the 50th state that didn't enact UA and and so but they have something comparable um so yeah is you vulnerable uh I don't think it is um you know it's been litigated a little bit but it's just very common sense basic law and nobody's ever talked about reforming or you know deleting it uh there was a federal law that almost preempted it called eign electronic signatures and global national Commerce Act but I was able to work with McCain who was then head of the Senate Commerce Committee and blly and others and we have a carve out for UA so it's not preempted so like this is like you can it's a loadbearing beam of the law so rely on it go forth and prosper okay thanks [Music]