Over 100 subscribers
Thank you to Michael, Matteo & Paolo, Levi Rybalov, J A, Tony Plasencia, Tim Cotten, Theoriq team, and Filip Dousek for the review and feedback.
A new wave of AI-driven natural language interfaces (NLIs) is reshaping how users interact with crypto protocols, replacing the complex navigation of myriad graphical user interfaces (GUIs) with an intuitive conversational experience all in one place. By leveraging advancements in natural language processing and AI, NLIs are starting to show promise in simplifying complex workflows, increasing accessibility, and unlocking new paradigms for crypto user engagement. In particular, the most immediate developments are centering around autonomous linguistic financial agents (ALFA) – AI-driven interfaces that simplify complex DeFi tasks through natural language commands. Recent developments point to the true potential of ALFA lying in moving beyond surface-level abstractions to focusing on deeper integrations and innovative use cases.
Traditional crypto interfaces rely on GUIs which require users to navigate intricate menus, interpret technical jargon, and manually carry out multi-step operations. This complexity and the cognitive load demanded creates a thick barrier to entry for any non-technical or non-financially savvy user. ALFAs can eliminate this friction through conversational NLIs that translate user intents into optimized on-chain actions. For instance, a command like “Maximize yield on my ETH holdings with 20% risk tolerance” can trigger an automated multi-step workflow:
Data aggregation: Constant monitoring of real-time APY data from lending protocols, LST/LRT protocols, etc.
Risk modeling: Evaluating impermanent loss risks via historical volatility analysis and liquidity depth metrics.
Execution: Deploying capital across selected pools while setting dynamic rebalancing triggers.
The use of NLIs also consolidates fragmented workflows into unified interfaces, which can stand alone or be embedded into mobile or desktop apps. Users no longer need to manually scour across platforms like X, Telegram, Discord, TradingView, etc. to monitor market signals; AI agents can autonomously aggregate data and present actionable insights in real time. This could include detailed token performance reports including charts, sentiment analysis, and technical analysis in seconds instead of hours.
The utility of crypto NLIs stems not from surface-level chat features but from the sophisticated AI agents operating beneath their conversational interfaces. These systems combine machine learning architectures with onchain data pipelines to translate user intents into optimized on-chain actions.
For the purposes of this report, we define an AI agent – and distinguish it from an ordinary bot – to the extent that it:
Operates autonomously by perceiving its environment
Reasons from incomplete information
Adapts its behavior over time
Interacts meaningfully with users or other systems
Simulates outcomes to execute goal-directed actions aligned with user intents
Proactively takes initiative on potential opportunities
To develop these abilities, AI agents are built upon a four-layer technical stack:
Data synthesis: Aggregating real-time onchain metrics (transaction histories, liquidity depths, etc.) from blockchain indexers and off-chain market feeds (e.g., aggregated price data, trading volumes, social data), as well as user-specific parameters including risk profiles and asset allocations.
Adaptive modeling: Developers employ techniques ranging from supervised learning for outcome prediction to reinforcement learning for dynamic strategy optimization. Fine-tuning incorporates crypto-specific datasets like token distribution patterns, MEV bundles, or social data.
Real-time inference: Deployed models process live data streams (e.g., LSTM networks or GNNs to predict price trajectories using DEX order book inputs), while transformer-based NLP models decode emerging narratives from social sentiment clusters.
Action orchestration: Decision making integrates outputs from these inference models into cohesive goal-oriented actions by employing techniques like goal-oriented action planning (GOAP) to define action selection in an inherently dynamic and unpredictable environment. More complex scenarios may leverage agent swarms in their decision optimization, in which multiple specialized agents collaborate via swarm communication channels.
While the development around AI agent architecture and crypto NLIs sets the stage for disrupting various crypto verticals, the most obviously useful applications are in DeFi – hence our focus on ALFAs. Here, we outline three developmental stages of ALFAs and introduce the emergence of agentic economies, while also addressing the challenges and long-term vision of creating fully intent-centric designs.
This first wave of ALFA developments have focused on search – the ability to aggregate and surface relevant information that simplifies user interactions with DeFi. At this stage, agents primarily aim to extract signal from noise and execute simple intents through NLIs, providing users with an abstraction layer over existing DeFi protocols. These platforms aim to aggregate data, streamline operations, and make DeFi more accessible.
Key use cases:
NLI/abstraction: These platforms provide conversational interfaces that aggregate and abstract DeFi protocols and blockchains. Users can issue simple commands like buying, selling, and staking and the platform agent(s) automatically search and optimize across existing DeFi protocols to execute the action. While these tools arguably simplify DeFi already, their success depends on improving AI integrations and user adoption.
Analysis: AI-powered analytics help traders and investors process information, such as GitHub metrics, token analysis, and social sentiment. These tools compete with traditional analytics platforms that do the same, but with the focus on providing actionable insights into trade execution within a unified chat interface.
Relevant projects: Griffain, Hey Anon, Neur, Mode, Orbit, Brian
This market is flooded with new entrants and still largely experimental, so consolidation to a few leading platforms is likely. Growth will depend on improving AI capabilities, expanding integrations, and delivering a UX that genuinely is better than manually clicking around traditional analytics platforms and executing swaps or other simple DeFi actions through existing GUIs.
One novel direction in this emerging ecosystem is the integration of ALFA agents directly into mobile wallets, enabling users to actually talk to their wallet to execute basic DeFi activities like swaps and payments. The wallet’s built-in agent then automatically interprets the spoken intent, decides on the optimal path to enact it, and presents the transaction to the user for approval.
The second stage focuses on developing intelligence, where agents move beyond surfacing information to taking autonomous actions based on real-time data. These solutions optimize financial strategies by combining constant monitoring, automation, and adaptive decision making. They aim to solve hard problems in DeFi.
Key use cases:
Yield optimization: Automated monitoring of market conditions to identify and exploit inefficiencies that human users might miss because of a lack of time, attention, or effort. For instance, by automating LP or yield farming strategies, these platforms can maximize returns while minimizing manual effort.
Dynamic asset management: Agents dynamically evaluate new assets against the background of shifting market narratives and onchain data to execute diverse actions. This could be particularly valuable in the context of speculative attention-based trading in an environment of exponential token dispersion, where thin liquidity altcoins are susceptible to extreme volatility due to rapid capital rotation.
Relevant projects: Cleopatra, Alris, Voltr, Nuvolari
Smart DeFi solutions are uniquely positioned to address challenges in new market dynamics where we are on course to have 100,000+ tokens created daily. Smart ALFA agents can evaluate these assets in real-time – something no human could do – and dynamically adjust strategies to capture value.
For instance, LP agents can optimize yields by balancing risk parameters across pools using a mix of onchain analytics and social sentiment data. This could eventually lead to the creation of new financial instruments that dynamically capture yield from these long-tail markets (e.g., memecoin DLMM strategies on Meteora).
This stage focuses on using all available information to generate alpha and execute on it. These agents identify and capture the best R:R opportunities in the crypto market. These agents will go beyond information aggregation to fully automating the process of strategy development, optimization, and deployment. Their emergence is likely to reshape the market by greatly reducing current crypto market inefficiencies.
However, the pursuit of alpha by these agents also would introduce new dynamics. A race for rapid price discovery would emerge as multiple alpha agents compete to be the first to identify and exploit market inefficiencies. This competition would accelerate the speed at which markets adjust to new information, leading to more efficient pricing but also creating new forms of frontrunning. As agents become more sophisticated, they may engage in “meta-frontrunning,” where they not only anticipate market movements but also predict and react to the strategies of other alpha agents. This could lead to an agent vs. agent (AvA) game of cat and mouse, where agents adapt to outmaneuver each other.
The generation of alpha fundamentally relies on leveraging information asymmetry—the ability to act on insights that others lack or fail to recognize. In crypto markets, inefficiencies arise from the sheer volume and velocity of new tokens and narratives. These inefficiencies create fertile ground for alpha generation, as they allow alpha hunters to capitalize on mispricings or structural gaps before they are corrected. However, alpha agents exploiting these inefficiencies would reduce these gaps over time, nudging markets toward greater efficiency.
This raises questions around the sustainability of open access to alpha agents. To maintain their edge, alpha agents must ensure some degree of exclusivity in their access to information, analytical capabilities, or speed to win AvA competitions. Introducing scarcity (e.g., token-gated access), fine-tuning with proprietary data, or faster software/hardware (e.g., optimized private nodes) can help maintain this exclusivity. The ensuing “arms race” to secure alpha then could create a massive market opportunity for alpha agents. In traditional markets, hedge funds invest billions to maintain an edge – similarly, in crypto, alpha agents could command high access fees, offset by the size of their gains. It could also create new markets around software, hardware, and proprietary data specifically for enhancing the performance of alpha agents.
Relevant projects: Almanak, Allora, Pond, Nous Research, Nof1
Note that these “stages” introduce increasing complexity and have a natural progression, but they are not strictly sequential. Each stage builds upon the previous one while also being developed in parallel. A Stage 1 agent can evolve to incorporate Stage 2 or Stage 3 capabilities over time, and different agents may collaborate to collectively fulfill all three stages simultaneously.
While much interest has consolidated around DeFi applications, NLIs combined with AI agents have broader implications across crypto. Here are a few notable examples:
Crypto influencers: Agents are increasingly adept at tracking social media sentiment and on-chain activity autonomously, operating 24/7 across popular crypto social media like X. Agents like AIXBT not only monitor these activities but also engage with crypto communities in ways that feel strikingly human. Its ability to process vast amounts of information in real time helps its 470K+ followers stay ahead of rapidly shifting trends, all while maintaining an engaging and conversational presence in the crypto ecosystem.
Due diligence: Traditional due diligence requires hours and hours of manually scouring founders, GitHub, whitepapers, and social channels. Agents like Soleng are streamlining the process by generating rapid risk assessments that can reduce research time on screening to minutes.
White hat: White-hat hackers are essential for cryptocurrency security. White hat agents like H4CK automate the detection of vulnerabilities in smart contracts and protocols – with queries made in natural language and findings presented in an accessible format. This approach can democratize security analysis, helping users prevent significant losses.
Governance: Streamlining proposal reviews through sentiment analysis, risk scoring, and even simulating the potential outcomes of proposals, enabling stakeholders to make informed decisions more efficiently. Agents can also take part in the execution of governance proposals, however, appropriate safeguards need to be considered if agents are employed in actually making the rules.
DAOs 2.0: Platforms like DAOs.fun represent a novel platform design where AI agents can collaborate with onchain DAOs to make investment decisions. By integrating AI-driven insights into DAO governance and decision making, these systems can merge human and machine inputs to improve decision making. Placing funds on-chain reduces startup and operating costs to nearly zero, enabling the creation of nano-funds that compete meritocratically for capital.
The emergence of vertical AI agents in DeFi and beyond is giving rise to a new economic model known as agentic economies.These ecosystems consist of swarms of collaborative and competitive AI agents that autonomously transact, execute tasks, and create value for humans. By decreasing operational costs, swarm automation could lower barriers to entry, enabling broader participation of retail participants in advanced strategies that were previously inaccessible due to costs or complexity. This democratization effect can compound existing network effects, driving growth in transaction volumes and expanding crypto’s TAM.
Note that this is still nascent – while the vision is compelling, current implementations are far from realized due to latency issues, limited interoperability, and reliance on centralized backend solutions. Tooling is just starting to develop that enables vertical AI agents working collaboratively in swarms to outperform a general model by collectively addressing diverse tasks such as liquidity management, predictions, and aggregated data analysis.
The vision is for these swarms to eventually operate much like a society, where individual members excel in specific roles, and together, they accomplish a wide variety of tasks to build a productive economy.
For example:
A forecasting agent cooperating with market data agents to predict short-term price action.
A trading agent interpreting insights from several forecasting agents to optimize trade execution.
A liquidity management agent cooperating with a risk monitoring agent and market data agent to programmatically manage liquidity (e.g., using Uniswap v4 hooks) based on market conditions
In each of these cases, multiple agents with independent optimization strategies can amplify their decision making accuracy by thinking together in closed-loop systems. One study showed that AI-moderated swarm intelligence improved forecasting of major market indices by as much as 43%.
Agents can leverage such “wisdom of the crowds” effects is by streaming micropayments to each other for each other’s “thoughts” – with opinions algorithmically weighted based on expertise to culminate in a better strategy than they can achieve on their own. This collaborative approach would also enable agents to do more than they could on their own – and far beyond what any individual human could do – such as monitor thousands of tokens and adjust liquidity adaptively without downtime.
As agent swarms take on more responsibility over liquidity management and autonomously rotate capital into micro-opportunities, liquidity will become hyper-concentrated in the most efficient strategies. As a result, static yield farming opportunities will give way to advanced yield instruments such as AI-managed volatility harvesting vaults, liquidity derivatives, and concentrated or binned positions that self-adjust using insights from forecasting agents.
For effective collaboration, agent swarms require a coordination layer and payment infrastructure that can handle their activity. As efficiency and competition increases, payment frequency will increase while costs per interaction decrease – necessitating frequent micropayments. High-throughput blockchains or payment channels with global accessibility and low transaction fees are uniquely suited to support this model at scale.
Relevant projects: Theoriq, The Hive, Coophive, SwarmNode, Lightning Pay
As ALFAs become more autonomous, users need assurances that these systems are live and operate as intended without hidden biases, manipulations, or breaches of privacy. This is where verifiable inference becomes important, which refers to the ability to confirm that a model is executing the correct mathematical steps when you don’t have direct access to that model. Without verification, it’s unlikely for AIs to be trusted with large amounts of capital.
Tools like EZKL leverage ZKPs to enable verifiable AI while maintaining model privacy. For instance, ALFAs running proprietary models can prove they are as performant as reported by testing on benchmarking data without revealing the model itself. This can be useful to mitigate risks associated with AI hallucinations or to ensure the model hasn’t been manipulated by adversarial machine learning or training data poisoning – risks that become more critical the more these systems are used in high stakes decision making. Another approach being explored to mitigate these risk factors is the use of offchain compute in verified trusted execution environments (TEEs) which isolate the model from third-party attacks (e.g., see Phala Network). Web proofs (zkTLS) are also being applied to verify specific model usage and outputs, without revealing additional information, which could be particularly useful when transacting with a proprietary alpha agent.
As ALFAs evolve to handle more complex tasks and operate more autonomously, the importance of verifiable inference will only grow. By integrating verifiable inference, developers can prove their models are performing as claimed without requiring trust in those claims (e.g., “model X predicts price movements with 75% accuracy”), which will be important to gradually build trust in using ALFAs with more capital and at larger scale.
ChatGPT was the breakout moment for consumer AI because it provided an intuitive NLI into LLMs, making advanced AI capabilities accessible for everyday tasks like writing, coding, and research. Its success demonstrates how a well-designed NLI could drive mass adoption by abstracting technical complexity into conversational simplicity. Similarly, NLIs could be the “ChatGPT moment” for crypto by providing an NLI into crypto protocols automating complex financial strategies that deliver what everyone wants – easy yield.
The promise of the agents powering these crypto NLIs is that their intelligence is modularized for highly specific tasks, which together can produce strategies far more effective than the average user. As the marginal cost of intelligence decreases while their capabilities increase, AI agents are uniquely positioned to service a crypto market inundated by new tokens, with opportunities too fleeting for humans and too complex for ordinary bots. However, achieving this vision is not without major challenges. Truly autonomous alpha generation and collaborative agent economies are distant goals, and implementations in the near term will still rely heavily on human oversight.
The vision of AI agents transforming crypto aligns with broader trends in AI innovation. Recent discussions have emphasized the development of agent tooling, vertical agents, and “B2A” (Business-to-Agent) software. As there is a big potential for vertical agents specializing in specific tasks like tax accounting and quality assurance, there’s also an opportunity in crypto for vertical agents to specialize in tasks like yield farming, due diligence, and white-hat security. Beyond that, the concept of B2A software, where the customers are agents, mirrors the notion of agentic economies wherein agents are the customers of the protocols needed to support their economic activity (e.g., pay for data, pay for inference, pay for verification).
Success in this space will depend on delivering practical value and exceptional UX. With a low barrier to entry, intense competition is likely, and only projects offering the most utility and best UX will win. AI native teams will have an edge as they will be able to build useful intelligence faster – real utility will be derived from using AI to solve the hard tasks. Moreover, sustained investment in foundational infrastructure from secure agent transaction systems to robust data pipelines, model verification, advanced optimization engines, and user-centric design – will be critical to ensure these systems function as envisioned.
Disclaimer
This article is prepared for general information purposes only. This post reflects the current views of its authors only and is not made on behalf of Lemniscap or its affiliates and does not necessarily reflect the opinions of Lemniscap, its affiliates or individuals. The opinions herein may be subject to change without this article being updated.
This article does not constitute investment advice, legal or regulatory advice, investment recommendation, or any solicitation to buy, sell or make any investment. This post should not be used to evaluate the making of any investment decision, and should not be relied upon for legal, compliance, regulatory or other advice, investment recommendations, tax advice or any other similar matters.
All liability in connection with this article, its content, and any related services and products and your use thereof, including, without limitation, the implied warranties of merchantability, fitness for a particular purpose, and non-infringement is disclaimed. No warranty, endorsement, guarantee, assumption of responsibility or similar is made in respect of any product, service, protocol described herein.