Backbone of a Verifiable & Interoperable Agentic Web I Yaoqi Jia at Pragma Denver 2025

[Music] I want to bring on yq to tell you on the same theme on how we actually have more verifiability with agents as we look at more and more of the agent universe so without further Ado let's welcome yq on stage goe yeah good afternoon everyone yeah glad to be here and um um as K te mentioned right today I will just uh share a little bit more on the recent research we did at this earlier and uh yeah basically um like like around one or two weeks ago we introduced this new platform called autonome uh as you can see right it's uh we we Tred to make it as this back bomb for this verifiable and also interoperable agented web uh we released the web paper we got a bunch of the feedback and uh later I will just show you it consists of a bunch of these components which are Super crucial uh not just um for this uh agent itself but I believe it's also for the entire sort of blockchain infrastructure in general for the next few years and we all know uh either in web two or web three right right now all the agents are quite U creats and uh we have all the different LL models like this open AI um enthropy Cloud uh deeps and also this recent Gro three right on top of that in web two we see bunch of these U amazing uh agents and also s like these Bots right uh we they can help us to do this traval planning and some of them can help us to do this payment and S similar to web 3 uh if you look at the recent development in this crypto space right we know that uh it's no longer just like sort of doing this a chat bot um similar to the chat gbt now we can also probably order pizza uh using some crypto agents very virtual or Eliza and why I mention about this because like we can imagine in the future apart from like other users either Web Two users or like sort of our crypto users right uh we we can interact with the agents but beyond that we like uh with this AGI right probably all the agents they will just talk to each other and basically to achieve some sort of goals together so in that case as you can see right it's no longer just a pure human in this picture we will also see a lot of these agents probably agent can also create agent so in that case right they need to talk to each other to achieve some goals very different executions of the tasks and S like to make everything autonomous and very secure and at the same time we can have very good responsive like sort of for this agent right the first component for this autonome platform we want to introduce is basically this P EVs and as you can see right we try to bring the heartbeat to these agents uh so far as we can see right the bunch of platforms um in this uh crypto space they create they basically help to host these agents in their environment mostly in the cloud right so in that case um as you can see right sometimes we don't know whether this agent can really work 247 or what's the s but for a lot of the crypto AI agents we all know that it's quite time sensitive for example if we are trying to ask the agent to do some def5 portfolio management or sometimes we want to do this swap on a snap some tokens on pump doine right it's super time sensitive basically if the agent is slower than like sort of the other agent and within like U like 10 millisecond right basically instead of sniping this kind of a new token right you are haunted by the other agents or by the other bots so in that case as you can see right um for different agent development um no matter is on Google cloud or AWS or different ones right or some decentralized hosting platform right we really need to make sure um they can help to make sure like sort of we have this very good um uh like like handling for this time sensitive uh like sort of operations here uh of course we Le this liquidation but for the other like sort of boxes also some other operation for example air drop claim and token swap um so these are all the very interesting application and very unique applications um in crypto and uh in web to we all know that if if we want to do transfer the best way like probably via strap or sometimes via Paypal right uh or if we go to bank right it's it's very slow but on crypto or sort of in blockchains right it's super fast we can do everything liquidation all the things on chain and uh uh easily so in that case as I mentioned right we have to have some ways to really monitor all the backends all this kind of engine uh agents like they can walk basically very a very timely manner and beyond all these time uh sensitive tasks right and sometimes we want to also guarantee this agent can really sort out provide these robust services and sometimes like before these agents are down due to various reason right CPU Ram disk we should also have the alert system for the agent or the owners itself so that's why as I mentioned right we introduce this PA EVS together with igen uh we have a different uh decentralized uh sort of set of operators to really monitor the status of the uh hostings of these agents basically like this a CPU RAM and also this cost and uh at the same time as I mentioned right it's about the less it's about the heartbe make sure all the agents actually when they are hosted on autonome we can make sure like they can provide the service with 99.9% or even higher s beyond that also uh like they can work 24/7 TR to make sure everything is is probably running and beyond that if something wrong right since EVS we can also slash these sort of like Stakes um uh belonging to this um uh s of agent and as you can see right apart from is a very uh decent monitoring and reporting for these agents we can also have better transparency since these operator can generate reports sometimes you can also check these online traces beond that of course better security compared to just single uh sort of Hosting instance for this um agent now we have a decentralized sort of cluster of nodes to really monitor the status of this um um of this agent and at the same time as I mentioned right um we typically finish all this heavy lifting and you don't really need to check the details for how to do uh how to write the EVS at the same time how this operator is running to support this uh EVs and and basically to check the status of the agents um at the same time right um as a developers if you want to really uh chip into like sort of the details of the operators and also this instan right typically we also provide these API and interfaces you can just query on top that um heartbeat and also this health check of these agents right we also want to make sure we have this verif verifiability of these um agents and and and as I mentioned right now we have plenty of these agents and one of these um um like sort of amazing agents we have in crypto is ax BT right but we all know that with ax BT it has very funny way and also intelligent way to reply to all the uh sort of commenters or some user right and then um we got lot of these concern and thoughts that whether the ax BT is really an agent or probably we have a bunch of students behind the scenes to reply like sort of these uh users on Twitter right and and so in that case it's very important to show that okay uh a like when we call it an agent is it a re real agent what's the LM model is using like sort of what's the uh what's the data source for this agent at the same time like uh if it's hosting a platform right uh the backend is it backend running and and at same time the reply on Twitter is it really from the agent or it's like from some other channels like some humans just tap and reply some funny content to the users so these are all the things we want to make sure like okay there's some verifiability um but in generaliz we can see right um if we look at this supply chain of this crypto AI typically it's not just about the agent hosting the front back is also about like sort of this data source so where we get the data and after that we where we do the training after training like how we do the inference and after inference like when we call these models right uh how we can really verify this back end of this agent of course when we talk about this verifiability we want to make sure we have this uh verifiability for each stage of this uh crypto air right um for example we need to use ZK proof for the data source make sure like okay uh there's nothing like sort of tampered by some third party authentic by the third party like sort of um uh information and beyond that also for training right we all want this decentralized training system we already see a bunch of these projects in the space doing the decentralized training but we all know that it's not that straightforward like it takes like tons of million dollars to have the gpus and uh in the data center to do training even for even we have the central way it takes a months of time and of course after deep seek probably will reduce the cost from from a tons of million to $5 million but still uh even in a centralized way it takes a time of course uh for inference it's much easier right you already see a bunch of these projects some of Partners like hyperbolic doing this inference in like like sort of the decentralized manner but here when we talk about agents in the current state typical we say it's about the hosting of the agent and typically we rely on some third party epis either from the open a gr or Lama um and we assume this data source uh for the infer uh for the for the inference is is is correct and then back to the point right as I mentioned so far the best thing we can do about the verifiability for the agents is mainly about the verifiability of this hosting for the agent backend um and what we are using for this very crucial compon component is a te plus ZK you can treat it something like okay uh we host the sort of this uh software uh for the back end of the agent uh in some te environment for example this Intel TDX sjx uh amdv and also there are some other open source ones for the res five and but in general as you can see right you can just treat it as a blackbox but the Black Box can guarantee that all the logic uh um sort of executed in that blood box is untempered by some third parties even the Hol machine in that case of course we need some extra uh proof uh to make sure like okay all the executions are basically guaranteed without any uh tempering right and beyond that the proof need to be verified on chain we also work closely with some uh project like this autom to do this uh uh proof verification on chain but in general you can see right the PA can provide this heartbeat and also health check for this agent on top of that we also have this second component uh with this trust environment and also Z proof to make sure this verifiability of these um agents and the third third thing which I feel like is very new and uh it's a little bit ahead of time in crypto space we call this cross agent routing system um like in short it's called car um because as I mentioned at the beginning right uh so right now we want see uh the agents either in web2 or crypto typical is sort of the user give the instructions like us give the instruction to the agent and the agent will we basically figure out the ways to to to to fulfill the the the instructions we give right but however as I mentioned right in the future if we have the AGI so agent probably can invoke the other agents to do and exercise some of these execution themselves um if you if you're using some uh deep research like sort of features from either Google Gemini or this like sort of this open ey 03 or like deeps right you can see that it's not that simple for the current uh agent right now it has multiple steps you can first search on the web uh regarding what kind of instructions or question you have right after that I can summarize these kind of uh materials searched from this um web after that it can do some deep analysis on top of this content but in the future right it can be much more complicated than that right now is still right straightforward like one agent doing everything but we know right in crypto probably we don't really just have one agent the one agent can help you to order pizza the other agents can help you to do this payment and of of course some agent can help you to do this defy portfolio management and some other ones doing this uh snapping on this pump Doan or on some like sort of this uh Unis swof right so in that case like if in the future right uh agents can be uh behave like this normal server or P2P machines in in the current stage and basically they need to find ways to discover each other and find I sort of the metadata so okay agent a is doing what and Agent B is u in charge of what so they can communicate with each other and basically perform some very complex like sort of operations and tasks and beyond that right there also some other features components we also need to make sure like um they can fulfill these tasks at same time they can get some payment um so that's why we have some very uh key modules for this uh cross agent routing protocol like registration adapters because like different models right they have different configurations and of course queing payment at the same time also different uh communication channels management but the good thing is like we already Implement most of the stuff if you are the agent hosted on the autonome uh by default you already have all these features of course for some third party or the external agent right if you want to communicate with the agents on autonome you can just follow some interfaces and uh call apis uh these are some C Snippets uh basically uh in the system and um it's a it's it's a it's already something uh we make it uh live and um and is like as I mentioned right um either you can just directly launch the agents on this platform or you can just uh call this uh apis or you just uh uh follow the SDK and write some code and then you can just interact with all agents either autonome or the other platforms uh via the vi this uh uh system um and uh and so far as I mentioned right there are three very important components we have all these um details in the whiteb paper uh if you have time you can just check out the our website and figure out this um the web paper and all the details uh the cool part that um like so we also have some mathematic formulas to prove that for this uh cross agent routing protocol um all the things is under this uh confidential environment oh another cool thing is about this confidentiality so all the data and message passed via this cross agent routing protocol right uh it's it's a conf is private um so it's super useful for especially defi application for example as a back to the example I use right if you are doing some snapping um pump doine or some like sort of ethereum evm based change right you don't want the the information to be laked to the other agents or users so in that case if you call the other agent right the channel should be confidential of course the user agent our communication Channel typical ver the TRS right but for the communication Channel across different agents it should also be very confidential um and beyond that be uh like all these three components already implemented for this autonome and we already done all the heavy lifting so all the products or features provided by earlier the good thing is like okay if you want to look at the details we have whiteb paper we have the mass proofs but at the same time we also finish all the heavy lifting so the staff we present to all these uh developers and users is just a no code platform and dashboard so uh you can just visit a website and um and as you can see right you can just choose a bunch of parameters and options and then you can launch a agent autonome uh which sort of like enjoy all the features I just mentioned and it's a pure no code and if you want uh you can just launch agent with the existing suppored Frameworks we have like for example coinbase uh Agent Kid and also uh Eliza and on top of that um together with uh East Global and card and other teams and previously we did hackathon I believe over like start of 100 developers already help to um sort of upload their Frameworks to this platform if you log to autonome I think right now apart from this U uh standard uh aging Frameworks as I show you uh you can also go to um the page and upload your own framework you can see this upload new framework right either you can f some of the existing ones or you can just create your brand new one and just upload the framework so later we can basically present your framework to the user developers and they can launch your agents and in the future probably some other features you can enjoy um as I mentioned right it's um um it's a still like sort of early for this crypto as space we hope in the future once we we have the AGI and uh we can have stand some standard protocols to help all the AG to communicate each other and uh at the same time uh as you know right for autonome we we try to present all the three M features uh the PA EVS for H harbit and uh this verifiability via the TZ proof and also this cross agent routing protocol car protocol to make sure all the message passed across different agents are under like sort of confidential environment and with this platform right we we also have this nood deployment platform just make sure everyone can can just U easily launch agent uh uh when you want uh we of course we have the event um on Friday and just to show more these demos from our partners there also some uh after party and happy hour but the the the the the thing I want to show you is really about this one just go to the website da. a.un uh this this uh this one is already live since last year you can just visit the website you can just click clear and then launch the agent yourself and I I recommend if you want to try to uh launch some rc20 or interact with unisoft you can try this agent key from this a coinbase team uh we're working closely with them uh with all their new features and then you can easily just tap okay let's uh launch a token yc2 token like it's Global and within like just few second they can finish it and then if you want to try this Twitter bot right like similar to a xbt and you can just launch the Eliza one uh link to your sort of Twitter account and U and just input some data source it can automatically sort of make the tweets and uh at the same time you can also fine tune uh this kind of agent to make sure everything uh is running following your instructions uh just have a try it's a very straightforward um all the things I printed at the beginning is already embeded into the system so you don't need to implement from scratch but of course um like most of you are very uh like sort of these Pro players in the development right you can also check out some some of the source code and probably you can build your own components on top of this autonome uh of course um if you encounter any questions just feel fre to join this uh TG group and uh and and uh for some existing like sort of Frameworks on the platform uh the authors typical in the group and if you want to extend the agent you running just feel free to tag us and our team will help you to uh uh extend like sort out the duration for your agents um yeah that's all the stuff and also all the new research uh we've done for the past uh few months earlier and uh if you have any question just feel free to reach out to me uh via this group or also our uh official Channel yeah really appreciate your time thank you [Music]