In crypto, we tend to write things off when money stops flowing. Whereas other industries would recognize the limitations presented by the thing they are working on, apply the Theory of Constraints, and optimize, crypto moves on to the next shiny thing.

Take AI research and development. It's wrought with constraints that often cascade into broader challenges - computational power constrains LLM training time, resulting in longer release cycles and exacerbating the limitations of dependent technologies. Where most industries vie for sustainability, crypto seems content with being a fashion that comes and goes.

"DeFAI" 🙄

Rather than identifying and exploiting the constraints presented by AI within the context of blockchain, cryptopreneurs jumped straight to the grift. This phenomenon stems from a fundamental tension in the crypto industry: brilliant technologists excited about the future collide with sieve-brained degens who will feign belief in anything promising a short-term win. Opportunists leverage these extremes to their advantage.

The pattern repeats endlessly. NFTs should have revolutionized gaming, but "GameFi" ruined their chance at bringing actual ownership to in-game items. ICOs should have revolutionized capital formation but instead became an indelible black mark on crypto's history. SocialFi had an opportunity to change how creators are rewarded online but naturally imploded due to greed and laziness. The metaverse was always stupid.

Perhaps immediately adding financially driven components to every product (most of which is vaporware) is not, in fact, a viable long-term strategy.

AI and crypto could be a perfect match if employed within the confines of reality and iterated upon as constraints are lifted and systems improved. However, mid-tier marketers LARPing as tech CEOs seem intent on torpedoing yet another significant advancement of our industry's many promises.

Understanding the difference between what we can do and what we are doing is paramount to the success of this vertical.

Agentic isn't a word.

Much like the metaverse - an ill-defined and utterly transparent façade over what essentially amounts to a poorly developed MMO - AI "agents" have a somewhat nebulous definition. If you ask ten people what an AI agent is, the three of them who know what you're talking about will give you different answers.

An "agent" should have complete and consistent awareness of its surroundings, giving it enough context to perform autonomous actions without intervention. Think about the agents in The Matrix. They have a primary directive (to eliminate anomalies within the program), but the path they take to do so is entirely novel and driven by a unique strategy. They even go so far as to find ways into the real world, possessing humans and machines to eliminate Neo.

In the real world (where we live), AI agents are chains of LLM prompts connected to backend functionality called "toolsets." These toolsets are akin to server routes that live among other tools in a graph. Using language, the LLMs determine which route to take and which tool to use based on their input. Like API endpoints, the routes can trigger system-level actions, such as swapping tokens. This essentially amounts to a new interface to a standard system - rather than pressing a button, you're typing words.

If that sounds lame, that's because it is - for now.

Adaptation.

Why add steps to bots and trading interfaces that worked fine with generic automation? Won't the novelty of launching tokens and managing a wallet in a social feed or via chat wear thin?

AI might improve blockchain interactions in many ways, but one stands out to me most prevalently. Adaptation.

While I can trivially write a script that automatically trades tokens based on various parameters, those parameters require constant monitoring as market dynamics shift. Running an effective trading strategy demands 24/7/365 attention - constantly adjusting settings, pulling levers, and pressing buttons. This situation is untenable, even impossible, for the majority of people. Herein lies my excitement surrounding AI.

I couldn't care less about AI agents becoming superstars on Twitter - I don't even care about humans becoming superstars on Twitter. What intrigues me is the potential of reinforcing each agent's learnings with actual data. Combining one's trading experience with the geyser of information online and onchain, along with the anecdotal, ever-growing knowledge base of an individual agent's performance, is an untapped gold mine.

Beyond situational adaptation, we can tie the agent's knowledge back to the Theory of Constraints. A well-trained agent will understand the limitations of its capabilities and integrated systems. If latency becomes a constraint due to mempool activity spikes, the agent can optimize execution by adjusting fees, tips, gas, and slippage in real time. Still problematic? It can "elevate the constraint" by swapping on an L2 with lower latency, perhaps even capturing arbitrage opportunities.

From this thinking, you can see how powerful crypto agents might become. But first, we must endure the phase of making them swap by saying "plz swap ser" to a chatbot. If this represents rock bottom, imagine what we'll achieve in 18 months, five years, a decade.