Infrastructure's AI Moment: The Race to Power Autonomous Agents | Shannon Wells & Molly Mackinlay

We are about to get started. Hey everyone, thanks for your patience. My name is Shannon. Um I'm an adviser to a bunch of web 3 and AI startups and I am really excited to dive into our topic um a uh infrastructures AI moment uh powering the race to power autonomous agents. Um so we are at a strange moment.

um AI is eating software, crypto is being reframed almost daily, and somehow these two things keep colliding. Um and the most interesting thing about this collision, I think, isn't that it's just about payments being what's going to um enable the breakout use case for crypto and AI. Um I think it's infrastructure. And we have the best person um I think in the building to talk to us about infrastructure um and decentralized infrastructure. Molly McKinley has been in the web 3 space since 2018.

Um, and before that she worked at Google um, on Chrome and mobile search. Um, she's been uh, an OSS contributor to IPFS since 2016 where she was the project lead. She led the engineering and research effort um, to launch the Filecoin Filecoin mainet in 2020. Um she launched the protocol labs venture studio uh and she's currently CEO of Phil Oz and leading the development of filecoin onchain cloud. So she's been building the foundational pieces of digital infrastructure for over a decade.

Um and now Molly you are watching the AI agent wave arrive um at those foundations. So let's get into it. Um there are roughly 10 million AI agents active today. Um, analysts expect that that to hit hundreds of millions within three years. These agents aren't browsing the web and clicking buttons.

They're transacting, allocating compute, storing state, calling APIs autonomously, continuously. That's a fundamentally different class of infrastructure consumer than a human developer. So why are agents specifically the forcing function for dependent and web 3 to find its moment today? What do autonomous systems actually need that's different from what human developers need? And where does traditional cloud AWS start to crack under that?

Awesome. So excited to be here. So excited to be chatting about this topic. I'm sure we're all like electrified by the momentum that AI is having in our lives. Like I I've been talking to a number of of developers I've worked with that have gone from a year ago, you know, working with claude code u maybe getting 10% lift to now like they don't even write code anymore.

They're just reviewing PRs as fast as they can from a whole team of of AI agents that are um making new changes to Solidity smart contracts. We've got our some of our s first uh vulnerabilities being committed by cloud code into solidity smart contracts as well. Um, so I think I think actually a h 100red million in three years is an underestimate. I would expect that we're at 100 million before the end of this year. Um, and probably even 10 million might be an underestimate of the number of agents that are already there but just aren't uh interconnected and interoperating um in the live uh internet or or accessible to to some of these estimates.

Um, I think it's a huge factor for all of humanity. Like we are riding a wave of creating a huge new workforce and portion of the economy. Um, and making sure that we build the the tools and infrastructure for that giant economic swath to use and operate on effectively is going to be a very big market. It already is. We already see like uh neoclouds uh building for AI training and inference workflows.

Um but as we start to think about agents themselves as autonomous actors um they have a set of priorities and a set of needs and requirements that don't mesh well with uh the traditional cloud or the legacy cloud um and work really well with some of the um areas that have been both selling points but sometimes hindrances for web 3 adoption. So that's why it's a really exciting moment for all of us who have been building infrastructure and tools in this space. Um, some of the things maybe as you're thinking from the the perspective of an AI agent where web 2 might get in your way. Um, one of those is just these clouds were designed from the ground up for humans. They're around human identity.

There's some individual's bank account on the back end. Um, they're connected to someone's email address. Uh and you know in a world in where we start to have larger and larger communities of agents interoperating and perpetuating themselves and paying for their own resources, clouds that are designed around a single human identity. and that humans may be liable for everything that agent is doing on the internet starts to become a significant concern for um those agents both in their own preservation but also having the um you know ability to to operate in online contexts autonomously. Um, you don't want to be dependent on an AWS API that might go down.

Literally the day that we launched the Falcoin onchain cloud um was a uh AWS and Cloudflare outage. Almost everyone's demos were down except for the folks that were building on a full resilient open cloud that didn't rec rely on these central points of failure. If you're an AI agent and your literal memory and ability to act in the world depends upon those single points of failure, you're pretty motivated to make sure that you have multiple different fallbacks and and ways to operate resiliency resiliently. And I think that's a a huge uh selling point for the the clouds and the the infrastructure tools that we've been building in web 3 is that there's multiple layers of fallbacks. there's very very very high uptime um because you have all of these different providers that are interoperating to uh come together to offer that service.

Um another aspect I I think itself has been a hindrance in web 3 for human adoption but is actually going to be a huge selling point uh for AI agents which is the fact that everything's kind of API CLI first. Um one of the the friction points for web 3 has been the ease of adoption for an enduser. They don't want to deal with their wallet interfaces. They don't want to have to call these different APIs. They don't want to click accept 17 times to bridge funds from like, you know, network A to network C.

But those are, you know, if they're just programmatic API calls, that's not so much a hurdle for an AI agent that is running on a super fast cycle time and is embedded in a API native context. Um, and so it a number of the the open APIs, the documentation, um, the open verifiability where an agent can see what's on the other side of the service that they're purchasing and verify it in real time. Um, that's actually a pretty big selling point for web 3 native services um and infrastructure tools that can meet agents without having to call up some uh sales provider, create a voice, create a phone number, create a fake human identity, create a credit card in order to purchase a service that really should just be an API call away. Um and so those are some of the the um building blocks that I think make this an exciting moment for web 3 to convert um in the AI agent market. Um obviously also where AI agents start harnessing um you know digital payments uh that's also an opportunity for our infraervices that are directly tied to a wallet address and uh a cryptocurrency payment.

um those are then accessible to AI agents in a way that you know purchasing an AWS bucket. You can't do that with USDC without uh some sort of human identity behind it. Um and so I think that's going to be an exciting movement.

Cool. So agents are giving web 3 a second wind uh and deepen specifically and you've been building in the decentralized infrastructure space now for 10 years. um what can we learn from the past um you know decentralized storage compute hosting um that we can keep in mind as we try and capitalize on the opportunity today?

Yeah, I think there's a few lessons that we can look at from the uh maybe some of the missteps or some of the the hurdles that the deepen ecosystem has overcome over the years. Uh a critical first one is focusing on supply versus demand. I think uh many deepen networks in building up this two-sided marketplace of service providers and service clients f focused first on deploying resources to build up larger pools of service providers. Falcoin did this as well. Um and and that has a hurdle because you're trying to emulate a client's need.

Um you're building up a large pool of available resources. Look at all of the the GPU um you know uh compute clouds right now. uh you have these large communities of participants and you're subsidizing their presence on your network without them doing useful revenue generating work on behalf of your ecosystem. Um and I think where we can focus in this moment is really closely mapping the inbound revenue to the work that's being done within the the infrastructure network. So, um I'm really bullish on, uh orienting the cryptoeconomics of networks and um making sure we're we're focusing on inbound client needs that are being paid for by a real user or client.

Some folks have done this via focusing on a single large proof of concept or a single client making sure they deliver against that you know LOI orou to a real paying customer and then they target the infrastructure buildout to meet that client's needs and then they scale horizontally from there. The other model is just you start with a a very clear value flow where 99% of the value a service provider in your ecosystem gets is the revenue that's flowing from a customer to that service provider so that there's constant alignment to meeting the needs of that provider within the ecosystem. So I think that's a key area where we've seen deepin networks kind of stray um in in building out more supply than they actually had demand to fill. Um, and also from the cryptoeconomic side, making sure there's a really clear model of how how clients are directing rewards and resources uh to those service providers based on quality of service and performance, not just based on like is there a checkbox somewhere in a proof that might or might not mean that I got my needs met as a client. Um, so I think that's one kind of critical uh journey.

I also think from a a deepen perspective um making sure that that we build things that are kind of fully endto-end being used. So looking at the whole user journey, not just the early hackathon PC. Um I think this will you know be those uh applications that are actually going end to end that are scaling that are becoming kind of like core building blocks that um are getting used. I I've started to see a little bit less focus on building out 10,000 builder communities and more focus on what are these like lighthouse applications that are scaling quickly that are gaining real usage that are driving real revenue. Um I think we'll see more of a a contraction and a focus even in the AI infrastructure age um into the the real applications products and use cases um that are you know converting and and actually uh scaling their adoption not just the number of examples that you can list in a uh a promo website.

Um, and so I think that's an exciting opportunity as well for teams that really want to build real products versus, you know, uh, mining different rewards and bounties at hackathons. Um, I also think it's a a really cool moment where we're using AI agents to build out those tools for themselves. Um, and that can create a a flywheel of new product ideas that are getting hacked on. It's crazy to me that we see so few of those really interesting new ideas coming out of the the large AI labs. I think they're just they're too busy doing their model training.

They're not doing the product experimentation that is available in the wider ecosystem. Things like, you know, Open Claw, Molt Book, those aren't, you know, coming from the large AI labs. That means there's a great opportunity for this ecosystem as well to build uh some of those new products as they can grow. Yeah, I want to tap into that because I I want to debate you on this point. Actually, I think that the AI labs have been paying attention to blockchain.

Um, just yesterday, Open AI, which actually acquired OpenClaw, um, after Anthropic sent the founder like a cease and desist letter or something. So major fail for anthropic, but Open AI and Paradigm Research, which folks in this community will know, published a joint research paper um with Tempo, which is a purpose-built layer 1 blockchain for stable coins, um on EVM bench, and it's a benchmark for how well agents can detect and patch high severity smart contract vulnerabilities. Um Anthropic also published a paper in December just on um how to benchmark for smart contract vulnerabilities. The use cases here are around finance. So I think we're seeing AI labs recognize the value prop around stable coins.

But how does the infrastructure space in web 3 kind of take that um take the playbook of stable coins to like actually start to get in the door of these AI research labs and of these AI startups because we're not going to win if we just keep talking to ourselves in web 3. We have to be going out to the broader market. So how do we how do we do that?

I think a couple ways like one there's going out to the broader market as in going out to the agent communities themselves. Um Salana was actually doing a really interesting thing. All of their websites now you append MD and it becomes like an agent readable um skills file for how you use that particular thing. It's good insights. If you make your documentation and your tools the most accessible, most well doumented, most open resource, then anyone who's using agents or the agents themselves that are deciding what tools to use to build out their their next workflow are going to default to the ones that solve their needs the best.

And so I think there's a a key kind of DevX opportunity there that that also resonates with where I think this ecosystem has historically catered toward developers and open documentation. So I think that's an area to lean into. Um I also think there's a a flag around like the global access side of things. There's just so much demand like hardware costs are rising massively right now. So any group that has uh existing hardware capacity, um helping open that up to uh ecosystems that literally can't buy more of the hardware they're looking for.

Um I think that's a a good opportunity for people to to monetize um and you know help groups scale that are getting priced out of Amazon or Google or traditional cloud operations.

Cool. One last spicy question. Brian Flynn yesterday wrote, "If your service can't be discovered by a machine, it doesn't exist to agents. Um, discovery has to be programmatic. A human still decides which tools an agent is allowed to use, but once the agent is running, purchasing decisions are pure optimization."

So, for onchain services specifically, what does that discovery layer actually look like?

Yeah, I mean, the 8004 community is doing a discovery layer around all of the different agents that exist. I think from an infrastructure perspective that's having open APIs and good documentation about the problems that you're solving. Making sure that you have good SEO equivalents so that those are uh top of mind and describe well the problems that they're solving so that when you you know ask chatbt what API should I use for decentralized storage you get the filecoin onchain cloud instead of getting hey you should just stick it all on AWS that'll that'll be fine that's not going to break. Um, and so I think that level of of access is a a key thing to focus on. Um, I I would love to see even more tools where we're kind of enabling this kind of agentto agent uh prioritization and and evaluation of tools.

Benchmarks are are a great component in there as well as we can um start seeing different tools ranking on benchmarks in terms of supporting specific use cases. Um, and so like three different opportunities there. Make sure that your agents uh themselves can access this information. Make sure that your documentation is truly accessible. Um make sure that you actually benchmarked end to end the use cases you're trying to solve.

And very quickly, what should folks know about Filecoin onchain cloud and what's coming?

It's really exciting. We just have deployed all of the smart contracts on mainet. It went to test net back in uh November uh of 2025. Um, and so folks that are excited to build fully autonomous onchain agents, um, or are thinking about the the use cases for their web3 applications to, um, be storing their data in places where agents can then interoperate with those data sets effectively without centralized intermediaries. Um, Falco onchain cloud runs entirely in smart contracts um, entirely with wallets and stable coins.

You can pay for storage and retrieval in real time on a global network of storage providers in the Falcoin ecosystem. Um, we have a lot of really nice SDKs. We have an MCP server. It's built for this exact community of builders. Um, and so we'd love to see people start integrating it into their workflows.

And if you are running cool things where your agents are programming against it, we have a whole set of RFS's requests for for startups built by agents. Um, and we're doing a hackathon right now uh as part of PL Genesis where you can build some of those cool ideas or whatever your heart desires. Um, and we'd love to see what you what you output. So, um, tell me what you're doing. Cool.

We are over time. Um I think we'll be around for questions if people want to chat with us after. Thanks very much.

Thank you everyone.