Beyond conversational interfaces, a primary usage of AI is the automation of long-range, high-autonomy tasks such as coding and research. The focus on autonomy has given rise to the popular characterization of agents as sovereign, anthropomorphic entities. In practice, agency is an emergent property of applications, not an inherent quality of the model. The LLM itself is stateless. Agentic orchestration is fundamentally an backend software that processes data, manages states, and calling function tailored for non-deterministic computation.
Despite the misconception, the industry has now entered a phase of building rails for agentic entities, including Agent-to-Agent (A2A) communication frameworks and on-chain primitives like Ethereum's ERC-8004. Many have fueled the vision for an agentic economy, with some platforms marketing about the volume of autonomous agents they support.
While the vision is compelling, some projects encounter practical limitations - their usability based on economic premises that relies on agents deployed by third-party developers. In Crypto x AI 2025, it’s not entirely clear how sustainable platform economy will take off and create network effects that incentive “agents”. We might end up with systems that fall short on performance and utility, failing to deliver effective automation.
This vulnerability to narrative-driven development is not unique to crypto. Real-world businesses, despite facing direct P&L pressures, are just as susceptible. Recent enterprise surveys paint a sobering picture of a widespread "pilot-to-production" gap. Fortune reported that 95% fail to hit their target performance, because these AI initiatives result in generic MVPs that don’t integrate into specific, pre-existing enterprise workflows. While the rest 5% that succeed do so through a disciplined focus on a single, well-defined use case.
As the sector march toward grounded financial use cases, many companies are effectively becoming fintech firms, and this shift requires more mature strategies. After the speculative hype of AI memecoins, crypto-fintech companies need to rewind their AI efforts to focus on what creates tangible value for customers. For example, an NFT marketplace could pivot as an AI-generated art platform with A2A trading, while it could simply optimize its search bar with natural language chat, helping users find what they want with better UX. A DeFi platform can build agent swarms to allocate vaults, but it might be more practical to transform existing data pipelines and allocation workflows as LLM tools so that human manager can analyze and execute more efficiently.
The alternative approaches, while less spectacular, often delivers more immediate and sustainable business impact. If the endgame is fully autonomous agents acting on the emerging A2A protocols, we must start by building the harnesses around, powering tangible automations in bespoke corner of the industry.
There is no one-size-fits-all AI solutions. All automation needs to be tailored to operational use cases in a B2B fashion. In tradFi, automations on credit scoring, personalized financial services, fraud detection, and robo-advisory services is projected to hit $42 billion by 2030. While B2B enterprise has never been a venture back-able story in crypto, since its growth margin seems despicable comparing to launching coins.
We launch coins on memes, and the opinionated character of agents who yaps and trades embodies good memes. Under this influence, developers’ mindset also stirs towards implementations stressing on integrated autonomy, which can lead to sub-optimal architectural traps. Crypto builders are several cycles behind from frontier AI in LLM engineering practices. The core of a AI system simply follows Model + Tools + Context, and a cost-effective solution must optimize each of these components with modularity.
We can prompt the model of being a “financial analyzer who uses Excel”, but it’s counter-productive to assume a “financial analyzer” that run continuously in isolation, because every call to the LLM is independent, with input constructed from scratch by the framework. To elaborate in an agentic loop, the system can run the financial analyzer workflow interacting with the user in the first 5 calls, but at call #6, it can hide the conversation history and simply prompt “process the following document with tool X” without context overloading. If the business use case involves mostly data processing and little intellectual judgement, the framework should be flexible enough to jump out of the loop, and call the LLM with independent processing command.
