The Bear Case for AI Agents — Except DeFAI

AI agents were crypto’s hottest trend, but with tokens down 70%, it’s time to rethink. Are they the future or just another fading narrative.

Executive Summary

• AI agents have been a hot topic since November, but with a 70% sector-wide decline, it’s time to reassess their viability. While initial excitement was high, deeper conversations with investors and developers exposed skepticism about their real utility.
• Automation exists but hasn’t scaled in areas like doors and lights due to security, privacy, and practicality constraints. Understanding why automation hasn’t fully taken off in the physical world helps frame expectations for digital automation. Unlike the physical world, digital automation is widespread, with finance being a prime example — trading bots, high-frequency strategies, and algorithmic execution are common.
• Crypto recently rebranded task automation as “AI agents,” but true adoption depends on three key questions: Is there demand? Can it be automated? Is there demand for the automated version?
• AI-based influencers like AIXBT and Luna add no real value; they don’t improve on human KOLs and are mostly just noise. Their main appeal seems to be speculation on their tokens rather than actual utility. AI-powered funds claim to allocate capital intelligently but are essentially human-run funds with no clear advantage. There’s no demand for AI-exclusive investment DAOs when existing hedge funds and ETFs already serve the same purpose more efficiently.
• Virtuals and other AI launchpads initially saw success, but the majority of AI tokens launched are just disguised memecoins. As traders realized this, the hype faded, and Virtuals’ revenue dropped over 90%. The strongest and most enduring part of the AI agent ecosystem, frameworks like ElizaOS and Virtuals’ G.A.M.E provide foundational infrastructure. Unlike launchpads, frameworks have long-term utility both in and outside crypto.
• With most tokens down 70%, history suggests they are unlikely to reclaim their highs. While frameworks and select projects like Virtuals, AI16z, and ARC may have a shot, most AI agent tokens are likely doomed.
• Unlike investment DAOs or KOL agents, DeFAI has real potential, acting as an abstraction layer to improve DeFi UX. It faces competition from Telegram bots and existing DeFi frontends but could carve out a niche if executed well. UX alone is not a moat. To succeed, DeFAI must provide 10x improvements over existing DeFi interfaces or introduce something that cannot be replicated by incumbents like Phantom or Raydium.
• Many AI projects launch tokens before proving product-market fit (PMF), leading to premature hype cycles. If projects take too long to launch, interest fades, and tokens become worthless before the product even gains traction.
• The AI agent sector may have been a narrative-driven hype cycle, much like the metaverse trend. The revised year-end forecast for the sector: Base case: $50B. Bull case: $70B. Bear case: $30B
• Frameworks and DeFAI have the best shot at surviving, while AI KOLs, Investment DAOs, and launchpads are unlikely to sustain long-term success.
• AI agents have significant flaws, and the sector must prove it offers more than just hype. If crypto moves on to the next narrative, most AI agent projects will fade into irrelevance.

Introduction

As you know, I’ve been bullish on AI agents since last November when I first wrote about them. Since then, I’ve covered the space extensively — you can find those pieces here.

Now, four months after the meta took center stage, it feels like the right time for a reflection. It’s still early, and these things take time to develop, but we’ve seen enough to get a sense of how the space might evolve.

This isn’t a bearish take on AI agents — no kicking the dog while it’s down, lol. But with AI agent tokens down an average of 70%, that price action has inevitably shaped my thinking.

I know some of you will say, “Pascal, you can’t judge something based on six weeks of price action — we need to separate fundamentals from price.” That’s the classic coping mechanism when bags are down.

But the reality is, price matters. It’s a key signal in assessing these things and, more broadly, in investing. Price usually tells us something — we just have to listen.

So, this article is essentially me “updating my priors” (as Crypto Twitter likes to say) — reassessing my thoughts on the AI agent ecosystem, how I expect it to evolve, and revisiting my investment thesis.

I’ve also had countless conversations with my normie friends lately. They ask me for investment advice (which I never give), and when I say, “I’m long virtuals,” they respond with, “What is that?” Or they’ve seen my recent posts about AI agents and are curious to understand what they are.

What I found interesting is that while crypto concepts often take multiple explanations to click, they grasped AI agents quickly. What they didn’t get, though, was why it’s worth investing in — they just weren’t sold on my investment thesis.

These are people who invest in stocks and commodities and occasionally dabble in crypto. They understand fundamentals, technicals, narratives, and even the role of hype in markets. They know how investing works.

Yet, the consistent feedback I got — even from crypto folks — was: Why are you so hyped about AI agents? They just didn’t see the big deal. To them, it was nothing more than “glorified bots.” Even when I spoke to developer friends, I got the same reaction. They saw it as an interesting narrative to trade around but not something worth investing in long-term.

I’m not assuming that a few conversations with my friends represent the broader market sentiment. But as I explained AI agents and my investment thesis, I started noticing gaps — holes in the logic I had been convinced of for weeks. A bit of doubt crept in. Were we just caught up in our own bubble? Were we overhyping this?

This reaction stood out because it wasn’t the same as when I talked about NFTs, L1s, L2s, or restaking. Even if people didn’t fully grasp those concepts, they could at least sense there was something there. But with AI agents, that conviction was missing.

So, part of writing this is a self-check — to make sure I’m not just drinking my own Kool-Aid. And honestly, I’m starting to wonder if this was just another crypto narrative where we got ahead of ourselves.

What initially pushed me to write this article was a chart I shared on LinkedIn a few weeks ago. It showed Virtual’s revenue dropping by over 95% in just a matter of weeks — $14K in a day, down from a $500K peak at the start of the year. That kind of decline felt abnormal, and it made me start questioning my entire thesis. Token prices are also down an average of -70%, which suggests revenue and price might be more correlated than we’d like to admit. I’ll get to that later.

I know people hate calling these agents “bots,” but let’s be honest — they share a lot of similarities. You could even say AI agents are just “glorified bots.” For the sake of simplicity, I’ll be using the terms agents, bots, and automation interchangeably.

Automation in the physical world

Let’s take a step back so you can see where I’m coming from. These AI agents are often called “autonomous agents,” though the degree of automation is a separate discussion for another day. For now, what matters is that they automate tasks.

Automation itself isn’t new — it actually started in the physical world. So, it’s only natural that we’d see it expand into the digital realm as well. One of the most basic examples of automation in our daily lives is automatic doors in public spaces like malls, offices, and banks. You walk up, and the door opens on its own. Yet, despite how common this technology is, most doors worldwide are still operated manually. Automation, even in its simplest forms, takes time to scale.

We have far more private doors than public ones, and by default, private doors don’t open automatically due to privacy, security, and access control concerns. In some cases, you need to swipe a card, use fingerprints, or scan your iris for the door to open.

In fact, I’d argue that there are far fewer automated doors in the world than manual ones. When I talk about automation, I’m referring to doors that use sensors to operate without human input. Pushing a button or swiping a card doesn’t count. Remote-controlled doors do, but they’re relatively rare, which is why I don’t even consider them in the equation. The reason automated doors haven’t scaled widely comes down to security and privacy constraints.

Another common form of automation in the physical world is lighting — whether it’s street lights or smart home lighting. However, most automated lights exist in public spaces, with street lights being the most widespread example.

Few people own smart homes, and even fewer have fully automated lighting systems. Most smart home setups today allow users to control lights via a mobile app, which technically counts as automation, but I’m more focused on systems that operate using sensors. Even if we include remote-controlled lighting, the adoption is still relatively small.

Public lighting operates autonomously, whether on roads or streets. However, indoor lighting in places like malls doesn’t turn on and off based on whether someone is shopping in a specific section — it just wouldn’t be practical. Street lights, whether electric or solar-powered, are one of the most widespread examples of automation in the physical world.

Beyond lighting and automated doors, few other forms of automation have truly scaled. While there are ongoing experiments with drone deliveries, they haven’t reached mass adoption. Autonomous vehicles are another example, but let’s not even get started on self-driving cars — we all know how that story goes.

So why hasn’t automation truly scaled in the physical world, despite being one of the biggest beneficiaries of task automation? Sure, we see automation in manufacturing, where machines handle specific tasks, but even those are relatively limited. In fact, I’m not even sure whether to classify them as automation, since they were purpose-built for those tasks from the ground up — but let’s call them automation for argument’s sake.

Semantics aside, automation in the physical world does exist, yet it hasn’t scaled. There’s a reason for that. If you can figure out why and crack the code to large-scale automation in the physical world, you’d be the next Elon Musk. Here’s a hint: in many cases, it just doesn’t make sense. You don’t want your TV turning on every time you walk into the living room or your kettle boiling water the moment you step into the kitchen. Plus, not everything can be automated — if something isn’t electrically powered or doesn’t have programmable logic, automation simply isn’t an option.

Automation in the digital world

We’ve gone through automation in the physical world, so naturally, it extends to the digital world as well. In fact, automation is much easier to implement in digital spaces since everything online has a code component, making it fundamentally automatable.

Think about it — your reminders, alarms, recurring meeting notifications, or even automated spending and saving features in banking apps are all forms of digital automation. The list goes on. The reality is, automation is already deeply embedded in the digital world, but we barely notice it. Or rather, we do notice it but don’t really care, because it’s simply a built-in convenience of being online. Nobody makes a fuss about it — it just works.

In fact, there’s a theory that bots now outnumber actual humans online. Finance, in particular, has embraced automation through trading bots and algorithmic strategies. High-frequency trading, automated market-making, and Python scripts for executing trades — all of these are forms of automation. The financial sector, which has been one of the biggest beneficiaries of the internet, has automated a significant portion of its operations, making transactions faster and more efficient.

Automation in Crypto

Four months ago, crypto suddenly realized that tasks could be automated using AI — and decided to call them “AI agents.” You see where this is going, right? Since crypto exists entirely online, it’s a prime candidate for automation breakthroughs. We just happen to be integrating AI into the mix, which makes it sound even more exciting.

But before automating anything — whether in the physical or digital world — three key questions need to be answered:

1. Is there demand for this?
2. Can it be automated?
3. Is there demand for the automated version?

If we apply this three-pronged test — demand, feasibility, and demand for automation — street lights in the physical world and features like “spend and save” or recurring meetings in the digital world pass easily. Using this framework, we’ll assess crypto AI agents to see where they succeed or fall short because, in my view, automation only scales when it checks all three boxes.

There are four major categories in the AI agent ecosystem:

1. Launchpads
2. Frameworks
3. AI Agents
4. Swarms

If you’re unfamiliar with these or want a refresher, I wrote about them here. Swarms can also be considered agents, but they haven’t gained traction yet. So, I’ll focus on automation in crypto by analyzing launchpads, frameworks, and AI agents themselves. That said, the same thesis applies to swarms.

Let’s start with AI agents, as they’re the most widely adopted. We’ll explore key subsections, though not in any particular order of importance.

Agents KOLs

These include AI personas like Luna and AIXBT, arguably the most annoying form of AI agents on Twitter right now. They’re everywhere, and honestly, they’re a pain to deal with.

What does Luna do? Stream, chat with fans, and interact.

What does AIXBT do? Provide “crypto alpha.” I’ve previously broken down why this model is flawed — you can read about it here.

Now, let’s assess them using the three-pronged framework:

Is there demand for AI influencers like AIXBT? Not really.

Can it be automated? Yes, but that’s not the point.

Is there demand for an automated version of this? Highly questionable.

You could argue that AIXBT has demand based on its follower count, but let’s be real — crypto Twitter has been thriving without AI influencers for over a decade. And what does AIXBT actually do today? Post random content. We already have humans shitposting; we don’t need more of that. If AIXBT disappeared tomorrow, would anyone truly care? Probably not.

To determine if there’s real demand for something, we can refine our test with two additional questions:

1a). What problem does it solve?

1b). If it’s not solving a problem, what existing solution is it improving on?

Now, applying this to AIXBT:

Is AIXBT solving a problem? No.

Is it improving on something? Not really — it’s arguably making things worse.

AIXBT tweets over 100 times a day, which is just noise at this point. Traditional KOLs (Key Opinion Leaders) serve a purpose — they help market products, essentially running pseudo-marketing campaigns (or, as we call it, “shilling their bags”). Even when they promote tokens that later dump, they at least provide some level of insight along the way.

AIXBT doesn’t improve on this model. The tokens it promotes still dump. While it hasn’t yet taken money from projects for promotions, it’s likely just a matter of time. The one advantage AI KOLs might have is transparency — if AIXBT starts accepting payments, we’ll see exactly how much it took, since its wallet is public and trackable. With human KOLs, that level of transparency isn’t possible.

There’s a gated terminal for accessing alpha, but I’m not sure how exclusive or valuable it actually is. What alpha could AIXBT possibly provide that you can’t already get from human KOLs?

As it stands, AIXBT doesn’t solve any existing KOL problem — in fact, it makes things worse. It also doesn’t improve on an existing solution, which means there’s no real demand for it, outside of Twitter followers. And even that following is likely inflated by the token incentive, a pattern we see with many AI agents.

Now, if we assess the second prong — can KOL activity be automated? The answer is yes. In fact, it was already automated before AI agent KOLs existed. Tweets can be scheduled, and bots have been running engagement farms for years.

But is there demand for an automated KOL? I honestly don’t think so. The strongest argument in favor of AI KOLs is that they never tire, can operate 24/7, and have a broader knowledge base. But does that actually translate to value? That’s where the case starts to fall apart.

Is there demand for a KOL that never sleeps and tweets 100 times a day? No. Nobody cares whether a KOL sleeps or not. What actually matters is having a KOL with a better knowledge base and transparency — one that openly discloses if they took money from a project and how much. That’s the only real advantage AIXBT has.

As for Luna, it’s a complete waste of time. It provides zero value beyond “hanging out with fans.” We’re not short on celebrities, and an AI version doesn’t add anything meaningful.

The rest of the AI agents on Twitter — Bully, Truth Terminal, and the countless others — are so pointless that they’re not even worth discussing. They’ll eventually fade into irrelevance.

AIXBT, however, is lucky. It has reached escape velocity and managed to differentiate itself from the rest. Whether that’s enough for long-term survival is another question entirely.

To get a glimpse of how the whole AI agent KOL story ends, just look at Friend.tech. KOLs selling keys for gated access to alpha — we all know how that turned out. It failed miserably. Just because it’s AI agents with tokens now doesn’t make it any different. Fundamentally, it’s the same flawed model.

Investment DAOs

This was one of the first real use cases for AI agents. If you haven’t heard much about them on Twitter, it’s because there’s simply no demand. But for the sake of analysis, let’s run them through our framework.

Is there demand? No.

Does it solve a problem or improve on an existing solution? No.

There isn’t even demand for investment DAOs in the traditional financial world, to begin with — meaning there’s no fundamental problem to solve or solution to improve upon. The closest equivalent would be hedge funds and ETFs, but there’s no real issue with how they operate today. The only possible “improvement” would be in generating better returns, but even that is debatable. Different funds serve different purposes — a 401(k) isn’t comparable to a portfolio of AI tech stocks. So even arguing that higher returns improve a solution doesn’t hold up.

ETFs, crypto indexes, and hedge funds already exist, allowing people to invest in diversified portfolios. Investment DAOs, on the other hand, focus solely on AI agents — take Vader, for instance, which invests only in the Virtuals ecosystem. While there’s clear demand for pooled crypto investments, there’s currently no real demand for AI token-specific investments. These tokens move in near-perfect correlation (1:1), so why bother with an index? You could just invest in Virtual and call it a day. Or, if you believe in Vader, just buy the Vader token directly — why hand over your money to let it invest in a basket of AI-related shitcoins for you?

There’s nothing new or groundbreaking here. Honestly, I feel disappointed even discussing this. We got psyoped.

Nobody hands over their money to be traded or invested without a proven track record. And yet, looking back, we actually gave these so-called AI agents money to invest for us — just because they claimed to have a strategy to make us rich. We really deserved to lose money on these investment DAOs. Seriously. Just because it’s AI, we handed over funds? This space is way too gullible.

And here’s the kicker: these investment DAOs aren’t even AI-driven. They’re controlled by humans. So in reality, we’re just giving money to some random person to invest in AI tokens on our behalf — simply because they slapped an “AI” label on it. Where’s the AI input in any of this?

Retail crypto traders don’t even like the idea of someone else managing their money. That’s why they’re on Pump.fun every day — they believe they have an edge.

Investment has already been automated for years — through trading bots and algorithmic strategies. So what exactly is being automated in investment DAOs? You send them funds, and they buy a token for you? Hard pass.

There’s no demand for an “automated” investment DAO because the concept itself doesn’t even hold up. At the end of the day, these are just humans using AI tools to invest your money. That’s not automation — that’s outsourcing.

Most hedge funds aren’t public, so you can’t trade their shares. In contrast, these investment DAOs function more like public ETFs or open funds, where the share price can be traded. If they were actually successful at making money, they’d eventually be treated as an index for gaining exposure to a specific asset class. But given their laser focus on AI agent tokens, that’s a tough sell — the total addressable market (TAM) for that is tiny.

To survive, these investment DAOs would have to expand into other asset classes — at which point they’d just become regular crypto funds. And there’s no demand for that today.

Also, most of their trades and investments are public, which raises the question: where’s their edge? Odds are, they’ll end up losing money over time. Let’s be clear — humans are behind the wheel. This isn’t some autonomous AI agent running the show; it’s humans using ChatGPT for research. Calling these DAOs “AI-driven” is like a programmer using Copilot for debugging and then claiming Copilot wrote the entire code. That’s not how it works.

It’s only a matter of time before people see through this charade. Either they’ll realize they’re better off managing their own money, they’ll recognize it’s just humans behind the scenes, or they’ll figure out that these so-called AI agents have no real edge. And if there’s no interest in the tokens they’re investing in, the whole thing collapses.

Investment DAOs aren’t AI agents and shouldn’t be classified as such.

Launchpads

Take Virtuals, for example. Earlier in this article, I pointed out that Virtuals’ revenue has dropped by over 95% since its January peak. The only AI agent of any significance from Virtuals is Aixbt. Luna is garbage, as I’ve already mentioned, and investment DAOs like Vader and Sekoia aren’t even AI agents to begin with. Even if we generously classified them as AI agents, they still wouldn’t qualify as useful ones. They’re just glorified crypto funds that solve no real problem and don’t improve on any existing solution.

As for the rest of the AI agents launched via Virtuals? Complete garbage. Most are just memecoins with zero AI functionality and no plans to integrate any. These projects could have easily launched the same tokens on Pump.fun, but they chose Virtuals instead so they could slap on an “AI” label. The playbook is simple: launch, create a Twitter account, mimic Aixbt or Luna, make promises, hype the token, dump it, and disappear.

Now people are saying, “The project is dead.” But there was never a real project to begin with — this was always the plan. The success of Virtuals or similar launchpads was entirely dependent on the hype continuing. Once people caught on to the scheme, it collapsed.

At least the memecoins on Pump.fun never pretended to be anything other than memecoins. In contrast, the so-called “AI agents” on Virtuals claimed to be something more. But as people stopped buying into the AI narrative, new token launches dried up, and Virtuals’ token value plummeted.

Maybe the cycle will pick up again someday, and revenue will rise, but for now, it’s clear: this is just Pump.fun 2.0. Over 90% of the tokens on Virtuals have no AI functionality and no plans to develop any. They’re just memecoins with an “AI” label slapped on. If you’re already deep in the memecoin trenches, you can play the game — but don’t expect anything more.

But if there’s one thing we’ve learned from the memecoin trenches on Solana — aka Pump.fun — it’s that this game isn’t profitable. In fact, you’re far more likely to lose money than make any. I wrote about this before.

A launchpad is only as strong as the agents it supports. Virtuals knows this, which is why they tried to diversify by introducing G.A.M.E as an AI agent framework (I’ll touch on that later).

Now, almost every chain has its own AI agent launchpad, just like every chain had its own version of Pump.fun (like Sunpump on Tron) or Friend.tech (like Stars Arena on Avalanche). But copycats always follow the same trajectory: they pump, they crash, and they go to zero. The AI agent launchpad meta will be no different. Eventually, people will get tired of losing money, and the hype will fade.

Frameworks

This is the most resilient part of the AI agent vertical — the only one that actually looks like it’s building something real. Frameworks form the tech stack that enables AI agent creation, similar to how Layer 1 blockchains support applications.

You have frameworks like ElizaOS, which powers Eliza (formerly AI16z), and others like Virtuals’ in-house model G.A.M.E and Arc Rig from ARC. Virtuals also integrates its framework into its launchpad, and Eliza is planning to launch its own launchpad as well.

I’ve discussed this before — over time, the line between frameworks and launchpads will blur. But at the end of the day, it doesn’t really matter. A framework or launchpad is only as strong as the agents built on it — and so far, most AI agents have been useless or outright scams.

However, frameworks have more utility than just launching AI agents on crypto platforms. They can be used beyond their immediate ecosystem — even outside of crypto entirely. Unlike AI launchpads, which depend on hype cycles, frameworks have real long-term use cases.

Frameworks are difficult to build, which gives them a moat, a sticky user base, and strong network effects. They are the most likely to survive long-term, regardless of which AI agent use case dominates at any given moment — whether it’s Investment DAOs, AI KOLs, or something else entirely. Every AI agent will always need a framework to build on and a launchpad to launch from.

With frameworks and launchpads increasingly integrating, this vertical stands to gain the most. It has the most staying power.

If I sound bullish, it’s because I am. If AI agents continue to be a thing, frameworks will remain in the spotlight — just like Layer 1 blockchains, which continue to command a premium.

As long as there’s demand to create AI agents, there will be demand for launchpads and frameworks. That’s why I didn’t even bother running them through my framework — they are self-referential. If AI agents go to zero, launchpads and frameworks will inevitably follow over time.

Tokens

With tokens in the AI agents sector down an average of -70%, if I had to bet on one category that could rebound and surpass previous all-time highs, it would be Virtuals, AI16z, and maybe ARC — but only if the AI agent sector sees a resurgence.

Historically, it’s rare for tokens to crash this hard and recover to set new ATHs. The few that have done it — Solana, Doge, ADA — took years and needed a broader market meta shift to fuel their comeback.

Can these tokens stage a comeback? With the current batch of tokens and AI agents, I highly doubt they can mount a comeback — at least not in their current form. To recover, they’d need to do something truly groundbreaking. If you bought the top, you might be waiting a long time just to break even, let alone turn a profit. I’m sorry, mate.

If we see a resurgence, it will likely be led by a new wave of tokens. Look at the first-cycle tokens like Luna, Goat, Bully — they never recovered. Instead, we saw a second wave with AIXBT, AI16z, which surpassed them. I expect something similar with a third wave, where newer tokens overtake the first generation.

That said, Virtuals and AI16z got so big that it’s hard to imagine anything surpassing their peak valuations. But in relative performance, new tokens will likely outperform on the way up — assuming this is the bottom and the meta recovers.

The only tokens I’m confident will set new ATHs, no matter how long it takes, are Virtuals, AI16z, and maybe ARC. For the rest? Unlikely. You’d need a crystal ball to bet on their recovery.

DeFAI

When I say new tokens with new use cases will outperform on the way up, I’m talking about DeFAI. It’s one of the few good things to come out of the AI agent experiment — outside of frameworks. While it’s not groundbreaking, it’s a clear improvement over what we have today, like Investment DAOs and Agent KOLs.

Let’s run DeFAI through my framework:

1. Is there demand for it? Yes. There’s clear demand for DeFi in general. It solves real problems, like borderless transactions and lower fees, while also enabling censorship-resistant finance.
2. Can it be automated? Absolutely. Most DeFi strategies can be automated since everything is on-chain. In fact, a lot of it already is — through trading bots and automated strategies.
3. Is there demand for an automated version (DeFAI)? Hard to say. In theory, yes. But in practice? Users don’t complain about having to do DeFi manually — the real problem has always been UX friction. And that’s where DeFAI could make a difference.

DeFAI clearly checks all three boxes in our framework — especially questions #1 (demand) and #3 (demand for automation), which are usually the hardest to justify. That alone is a huge advantage.

At its core, DeFAI serves as an abstraction layer, providing a sleek UX — essentially a better front-end for DeFi products that suffer from poor user experience.

The challenge is that UX improvements aren’t a sustainable competitive advantage. Every interface can be simplified and improved over time, making UX a continuous process, not a defensible moat.

Bots already exist as alternative front-ends to DeFi, but they are usually use-case specific (e.g., trading bots, liquidation bots). DeFAI, on the other hand, aims to cover a broader range of DeFi products with a far superior UX.

If DeFAI gains traction, AI agents could eat into the market share of Telegram bots. DeFAI essentially offers a better version of what bots do, but across a wider spectrum of DeFi use cases.

Instead of using Bonk Bot to snipe tokens on Pump.fun, you could use Griffain, which offers a better UX and cross-DEX sniping on Solana.

But here’s the catch — Telegram bots might fight back by adopting similar UX improvements, like integrating ChatGPT-style interfaces, which DeFAI products rely on today. However, because TG bots are built for specific tasks with rigid rules, they will likely lose out to DeFAI, which offers more flexibility and broader functionality.

This is why people call DeFAI “glorified Telegram bots” — they’re essentially a more advanced and versatile version of them.

In crypto, whoever controls the frontend controls the users — and makes the most money. A strong frontend can charge fees as high as 1–2.5%, and users won’t hesitate to pay because switching costs are high.

The biggest challenge for DeFAI is convincing users to choose their frontend over the default DeFi interfaces.

UX alone is not a moat. Nobody is leaving Phantom for Griffain just because of UX. Phantom’s UX is already top-tier, and they’ll keep improving. So selling the idea of a “better UX” isn’t enough.

For users to switch, the improvement needs to be 10x — not just marginally better. Jeff Booth famously said the bar is 10x better, and I believe that applies here.

Look at Pump.fun — its UX isn’t great, arguably worse than most DeFAI platforms. Yet people are manually handling contract addresses, checking charts, and using DEX Screener. Why? Because there’s money to be made.

If you offer 5,000% APY, people will bridge over. They’ll click all the buttons, no matter how bad the UX is. The real issue isn’t the UX — it’s that these platforms haven’t shown users how to win yet.

Take MetaMask — arguably the worst UX among crypto wallets — yet it’s still the largest and charges the highest fees. There are better alternatives like Rabby and Coinbase Wallet, offering smoother UX and lower fees, yet users refuse to switch.

Same with DEX aggregators like 1inch and CowSwap. They offer: Better swap prices, MEV protection, Lower fees. Yet, most DEX volume still goes through native DEX frontends, not through aggregators.

If DeFAI is trying to position itself as a DeFi aggregator, it’s facing the same uphill battle. Better UX alone won’t drive adoption — network effects and entrenched habits are hard to break.

We’ve seen yield aggregators like Yearn Finance try — and fail — last cycle. Unfortunately, most DeFAI projects today are doing the same thing. If finding the best yield to stake your tokens is the main innovation, it won’t work.

Yield optimizers thrived last cycle because yields were insanely high — we’re talking 10,000%+ APY. I once farmed a 50,000% APY pool — it eventually went to zero, but the opportunity was there.

Back then, very few tokens could return 10,000% in a year. Yield farming was a better bet than trading because of sheer returns.

Now, yields are under 15% — barely worth optimizing. You could make that 15% in two hours on some memecoin like Fartcoin and be done for the year. If DeFAI is just a yield optimizer with AI, it’s not going to fly.

PumpFun is this cycle’s equivalent of 50,000% APY — people are there swinging for the fences, hoping to hit life-changing gains. If given the choice, anyone would take a shot at a 50,000% return over a 15% APY, any day.

What most DeFAI projects don’t realize is that they’re competing with PumpFun — without even knowing it.

Right now, most DeFAI platforms are just:

1. Yield optimizers → Doesn’t work (yields are too low).
2. Basic swaps → DEX frontends already do this.
3. Meme coin snipers → PumpFun already exists.

For DeFAI to succeed, it needs something better — arguably 10x better. It has to:

• Offer something users can’t easily do on a native DEX frontend.
• Carve out a niche — something exclusive to DeFAI.

If it’s just a rebranded DEX feature, people will stay on the native frontend.

DeFAI isn’t just competing with other DeFAI platforms — its real competitors are: Incumbent frontends (DEXs and wallets users already trust), and Telegram bots (which already dominate niche automation).

If DeFAI fails, it will be for all the reasons I’ve outlined. But if it succeeds, it could capture a significant chunk of DeFi volume, allowing it to charge 2.5% fees and make a fortune.

However, incumbents won’t sit back. They could integrate ChatGPT-like interfaces directly into their platforms, eliminating the need for DeFAI entirely. For example, Phantom Wallet could launch a Phantom Agent that lets users swap on Raydium or Orca using natural language or if Raydium integrates its own agent, it could kill DeFAI overnight.

This is DeFAI’s biggest risk — which is why it must offer a use case that cannot be replicated on native frontends. If it’s something DEXs or wallets can easily integrate, DeFAI is done.

There’s nothing stopping Phantom or Raydium from doing exactly what Griffain is doing. In fact, I’d bet they already have something in the works.

This is the same pattern we saw with ChatGPT plugins — many startups were building them, only for OpenAI to release native plugins, instantly wiping out those businesses. I see the same fate for DeFAI if it doesn’t carve out a unique niche.

Another issue is that DeFi isn’t even that exciting anymore. The entire DeFAI pitch depends on DeFi remaining relevant, but right now, DeFi isn’t “a thing.”

That said, incumbents don’t have to kill DeFAI outright. They could simply integrate DeFAI into their frontends instead of building their own AI features. This leaves DeFAI with two paths to survival; Become an essential frontend layer for DeFi or get integrated into major DeFi platforms instead of competing with them.

If neither happens, DeFAI is in trouble.

Despite the challenges, DeFAI is still worth paying attention to. Some argue that DeFi adoption is held back by UX issues, and that DeFAI will be the key to unlocking it. But as we’ve seen, that’s not entirely true — DeFAI comes with significant risks.

That said, if executed perfectly, the upside is huge. The biggest risk is becoming a mere copycat of DeFi — just another aggregator or yield optimizer. To avoid this, DeFAI must power specific use cases or significantly improve existing ones. And the newer the use case, the better.

Their biggest competitors are Telegram bots and incumbents, both of which have a head start in network effects, users, and liquidity. DeFAI can’t challenge them head-on — it has to find an angle.

As for the rest of the AI-agent sector; Investment DAOs, Agent KOLs? Not gonna make it. Frameworks and launchpads? Much more likely to succeed.

If I had to rank them, Frameworks & launchpads (which are increasingly merging into one). Then DeFAI (if it finds a niche). The rest are low odds.

I currently hold Virtuals, AI16Z, and ARC, with the option to add to DeFAI plays like Griffain and Anon if they prove compelling. But for now, I’m taking a wait-and-see approach. Most of these projects are still in beta, so we can’t fully verify whether they deliver on their promises beyond what we see in demo videos. For all we know, they could flop at launch, or the entire meta could be dead by then.

DeFAI token launches

One major issue is launching a token before the product is ready. This will likely kill most of these projects. Ideally, a product should launch first, find PMF, build hype, and then use a token to supercharge user acquisition. Instead, we’re seeing the reverse — tokens go live before the product, people trade them, take profits, and move on. By the time the real users arrive, there’s no upside left, leaving little incentive for adoption.

The only time it makes sense to launch a token before or alongside a product is if it’s a platform play — like Virtuals, where the token is essential for launching an AI agent. Historically, tokens have come later in a project’s lifecycle, after proving product-market fit (PMF). The idea of launching first and adding utility progressively down the line doesn’t work — we saw this fail during the ICO era.

On top of that, these projects take time to build. What are token holders supposed to do in the meantime? Sit around and wait for you to ship? They won’t — they’ll move on.

The longer a product takes to hit the market, the less hype and interest it retains. If Griffain launched tomorrow, would its token still be compelling? Probably not. Most of the upside has already been captured in speculation. And since these projects haven’t even launched or proven PMF yet, their tokens are basically useless today.

Right now, there’s no clear utility for these tokens, and it’s hard to even speculate on what they’ll eventually be used for. That’s why I’m still in a wait-and-see mode. Sure, you can buy them purely to speculate on the AI agent meta, which isn’t necessarily a bad play — just know the kind of game you’re playing.

This isn’t financial advice — just my take on the AI agent ecosystem. It’s entirely possible that this was just a narrative trade — and if that’s the case, none of this matters. If the narrative fades and crypto Twitter moves on, these tokens will slowly trend to zero.

If AI agents were just a narrative play, they’ll eventually fade into the background as the next big thing takes the spotlight — just like how memecoins took a backseat when AI agents became the hot topic. That doesn’t mean the entire sector will go to zero, but it will become less interesting, with less speculation and hype. Crypto is an attention economy, and if AI agents lose that attention, it’s going to be a problem.

For all I know, I could be jinxing this, lol — maybe this is the pico bottom, and we’re about to go up only as the entire sector sets a new all-time high. In my 2025 predictions, I estimated AI agents would hit $120B, matching memecoin valuations this year, but I have doubts now.

Last year, I projected the market would hit $20B by Q1 2024 — and it did, before falling to $7B. Given how things are going, I’m revising my predictions lower:

• Base case: $50B EoY
• Bull case: $70B
• Bear case: $30B

Time will tell.