The Intelligent Evolution of DeFi: From Automation to AgentFi

Author: Jacob Zhao Source: mirror Translation: Shan Ouba, Golden Finance Among the various segments of the current crypto industry, stablecoin payments and DeFi applications, two proven sectors with real-world demand and long-term value, continue to demonstrate strong growth potential. Simultaneously, the booming development of AI agents is becoming the de facto user interface for the AI industry, gradually assuming a core role in the interaction between AI and users. Amid the convergence of crypto and AI, and particularly the exploration of how AI can benefit crypto applications, the current focus is on three typical application scenarios: 1. Conversational Agents, including chatbots, virtual companions, and AI assistants. While many are still "wrappers" of general-purpose large-scale models, their low development threshold, natural interactions, and token incentives make them among the first to reach users and gain traction.

Second, Information-Integrating Agents

This type of agent primarily handles the integration and interpretation of on-chain and off-chain data. Projects such as Kaito and AIXBT have achieved initial success in web-layer information aggregation, but on-chain data integration is still in its exploratory stages, with no clear market leader yet.

Third, strategy execution intelligent agents

This type of agent is centered around stablecoin payments and DeFi strategy automation, and has two main directions:

1. Agent Payment

2. DeFAI (AI Agent in Decentralized Finance)

They are deeply embedded in on-chain transaction and asset management logic, and have the potential to move from "hype" to a sustainable, efficiency-driven financial automation infrastructure.

This article focuses on the evolutionary path of the integration of DeFi and AI, sorting out the stage transition from automation to intelligent agents, and deeply analyzing the infrastructure, application space and core challenges behind strategy-executing intelligent agents. The evolution of DeFi intelligence can be divided into three gradual stages: 1. Automated Infrastructure This stage is based on rule-based triggers and executes preset operations such as arbitrage, asset rebalancing, and stop-loss. The system cannot generate strategies on its own and does not have autonomy. Its execution logic is essentially "mechanical." 2. Intent-Driven Copilot: Introduces intent recognition and semantic parsing capabilities. Users can input commands using natural language, and the system is responsible for understanding and decomposing the intended action, suggesting an executable path. However, this stage still relies on user confirmation and cannot independently complete the decision-making and execution loop. 3. AgentFi (Intelligent Agent Finance): Achieves a complete intelligent closed loop: Perception → Reasoning/Strategy Generation → On-chain Execution → Continuous Evolution. AgentFi represents a truly autonomous on-chain intelligent agent with adaptive learning capabilities, capable of continuously optimizing its decision-making logic based on experience and achieving self-evolution. To determine whether a project is truly qualified to become AgentFi, it must meet at least three of the following five core criteria:

1. Autonomous perception of on-chain status/market signals (not static input, but real-time monitoring)

2. Strategy generation and formulation capabilities (not just preset strategies, but also the ability to autonomously create situational action plans)

3. left;">Autonomous on-chain execution (complex operations like swaps/loans/staking can be performed without human intervention)

4. Persistent state and evolutionary capabilities (agents have a lifecycle, can run continuously, and adjust their behavior based on feedback)

5. Agent-native architecture (dedicated agent SDK, specialized execution environment, and intent routing or execution middleware designed to support autonomous agents)

In other words: automated trading ≠ Copilot, and certainly ≠ AgentFi.

• Automated trading is merely a rule-based trigger system;

• The co-pilot can interpret user intent and provide actionable suggestions, but still relies on human input;

True AgentFi refers to an agent with perception, reasoning, and autonomous on-chain execution capabilities, capable of closing the strategy loop and evolving over time without human intervention.

DeFi use case applicability analysis:

In decentralized finance (DeFi), core applications are generally divided into two categories:

• Token capital markets: asset circulation and exchange

• Fixed income: financial strategies for generating income

We believe that these two types of products have significant differences in their compatibility with smart execution:

1. Token capital markets: asset circulation and trading scenarios

These involve atomic interactions, such as swap transactions, cross-chain bridging, and fiat on- and off-ramps. Their core feature is "intention-driven + single-step atomic execution" and does not involve income strategies, state persistence, or evolutionary logic. Therefore, they are more suitable for intent-centric Copilot and do not constitute AgentFi. Due to their low technical barriers and simple interactions, most current DeFAI projects fall into this category and do not form a closed-loop AgentFi system. However, more advanced swap strategies—such as cross-asset arbitrage, perpetual LP hedging, and leveraged rebalancing—may require the capabilities of AI agents, although such implementations are still in the early stages of exploration. 2. Income-Based Financial Scenarios: Income-based financial scenarios are a natural fit with AgentFi's "closed-loop strategy + autonomous execution" model, characterized by clear return targets, complex strategy combinations, and dynamic state management. Its key features include:

• Quantifiable return targets (APR/APY) enable agents to build optimization capabilities;

• A broad strategy design space involving multi-asset, multi-term, multi-platform, and multi-step interactions;

Frequent management and real-time adjustments make it very suitable for on-chain agent execution and maintenance. Due to many factors, such as the duration of earnings, frequency of fluctuations, on-chain data complexity, difficulty of cross-protocol integration, and compliance restrictions, different earning scenarios vary significantly in terms of AgentFi compatibility and engineering feasibility. We recommend prioritizing the following scenarios: High-priority implementation scenarios: Lending: Interest rate fluctuations are easy to track, and execution logic is standardized—ideal for lightweight agents. Yield Mining: Liquidity pools are highly dynamic, with diverse strategy combinations and high yield volatility. AgentFi can significantly improve annualized yield (APY) and interaction efficiency, but engineering implementation is challenging. Mid- to Long-Term Exploration Directions: Pendle Yield Trading: With a clear timeline and yield curve, it's suitable for agents managing rollovers and inter-pool arbitrage. Funding Rate Arbitrage: Theoretical returns are attractive, but challenges with cross-market execution and off-chain interactions require resolution, resulting in high engineering complexity. LRT (Liquid Restaking Token) Dynamic Portfolio Management: Static staking isn't suitable; potential lies in combining LRT with limited partners, lending, and automated rebalancing strategies. RWA multi-asset portfolio management: Difficult to implement in the short term. Agents can assist with portfolio optimization and maturity planning. Intelligent DeFi Scenario: 1. Automation Tools: Rule Triggers and Conditional Execution Gelato was one of the earliest DeFi automation infrastructures, supporting conditional task execution for protocols like Aave and Reflexer. It later transitioned to a Rollup-as-a-Service provider. Currently, the primary focus of on-chain automation has shifted to DeFi asset management platforms (such as DeFi Saver and Instadapp), which offer standardized automation modules such as limit orders, liquidation protection, automatic rebalancing, DCA, and grid strategies. Some of the more advanced DeFi automation platforms include: Mimic.fi – https://www.mimic.fi/ An on-chain automation platform for DeFi developers and projects, it supports programmable automation across chains such as Arbitrum, Base, and Optimism. Its architecture encompasses planning (task and trigger definition), execution (intent broadcasting and competitive execution), and security (triple verification and risk control). It is currently developed based on the SDK and is in the early stages of deployment. AFI Protocol – https://www.afiprotocol.ai/ An algorithmically driven proxy execution network that supports 24/7 non-custodial automation. It addresses the issues of fragmented execution, high entry barriers to strategy adoption, and poor risk management in DeFi. It provides programmable policies, permission controls, SDK tools, and a native yield-generating stablecoin, afiUSD. It is currently undergoing internal testing at Sonic Labs and is not yet available to the public or retail investors. 2. Intent-Centered Copilot: Intent Expression and Execution Recommendations The DeFAI narratives that emerged in late 2024 (excluding speculative meme-token projects) mostly fall into the category of intent-centric Copilots—users express their intent in natural language, and the system recommends actions or performs basic on-chain operations. Their core capabilities remain at the "intent recognition + co-pilot assisted execution" stage, lacking a closed-loop strategy and continuous optimization. Due to current limitations in semantic understanding, cross-protocol interaction, and responsive feedback, most user experiences are subpar and functionality is limited. Notable projects include: HeyElsa – https://app.heyelsa.ai/ HeyElsa is an AI co-pilot for Web3, allowing users to execute trades, bridges, NFT purchases, stop-loss orders, and even create Zora tokens using natural language. A powerful conversational cryptocurrency assistant for beginners, super advanced users, and seasoned traders, it's fully operational and live on over 10 blockchains. With 1 million daily trades, 3,000 to 5,000 daily active users, and integrated yield optimization strategies and automated intent execution, HeyElsa provides a solid foundation for AgentFi. Bankr – https://bankr.bot/ Bankr is an intent-based trading assistant that integrates AI, DeFi, and social interaction. Users can issue natural language commands through X or its dedicated terminal to execute swaps, limit orders, bridges, token issuance, NFT minting, and more on Base, Solana, Polygon, and the Ethereum mainnet. Bankr has built a complete "intent → compile → execute" process, focusing on a minimalist user interface and seamless integration with social platforms. It uses token incentives and revenue sharing to drive growth. Griffain – https://griffain.com/ A versatile AI agent platform built on Solana. Users can interact with Griffain Copilot through natural language to query assets, trade NFTs, manage limited partnerships, and more. The platform supports multiple agent modules and encourages community-building. Built on the Anchor Framework and integrated with Jupiter and TensorFlow, it emphasizes mobile compatibility and composability. Currently, it supports over 10 core agent modules and boasts powerful execution capabilities. Symphony – https://www.symphony.io/ – is an execution infrastructure built for AI agents, providing a full-stack system encompassing intent modeling, intelligent route discovery, RFQ execution, and account abstraction. Their conversational assistant, Sympson, provides real-time market data and strategy recommendations, but does not yet offer full on-chain execution capabilities. Symphony provides foundational components for AgentFi and is committed to supporting collaborative agent execution and cross-chain operations in the future. HeyAnon – https://heyanon.ai/ A DeFAI platform that integrates intent interaction, on-chain execution, and intelligent analysis. It supports multi-chain deployment (Ethereum, Base, Solana, etc.) and cross-chain bridges (LayerZero, deBridge). Users can use natural language to exchange, lend, and stake, as well as analyze market sentiment and on-chain dynamics. Despite the attention of founder Sesta, DeFAI is still in the Copilot phase, with incomplete strategies and execution intelligence. Its long-term viability is still under testing. The above scoring system is primarily based on product usability, user experience, and the feasibility of the public roadmap at the time of the author's review. It is highly subjective. Please note that this evaluation does not include any code security audits and should not be considered investment advice. 3. AgentFi: Strategy Closure and Autonomous Execution We believe that compared to the intent-centric Copilot, AgentFi represents a more advanced paradigm in the intelligent evolution of DeFi. These agents possess independent revenue strategies and on-chain autonomous execution capabilities, significantly improving user execution efficiency and capital utilization. In 2025, we are delighted to see a growing number of AgentFi projects either launched or in development, primarily focusing on lending and liquidity mining. Notable projects include Giza ARMA, Theoriq AlphaSwarm, Almanak, Brahma, the Olas agent series, and many more. ARMA – https://arma.xyz/ ARMA is a smart agent product launched by Giza, designed to optimize cross-protocol stablecoin yields. Deployed on Base, ARMA supports major lending protocols such as Aave, Morpho, Compound, and Moonwell, offering key features such as cross-protocol rebalancing, automatic compounding, and smart asset switching. The ARMA strategy system monitors stablecoin annualized interest rates (APRs), transaction costs, and yield spreads in real time, and autonomously adjusts its allocation to achieve returns significantly higher than static holdings. Its modular architecture includes: smart accounts, session keys, core agent logic, protocol adapters, risk management, and accounting modules. Together, these modules ensure non-custodial, secure, and efficient automation. ARMA is now fully launched and rapidly iterating, making it one of the most practical AgentFi products in the field of DeFi yield automation. Theoriq Alpha Protocol – https://www.theoriq.ai/ Theoriq Alpha Protocol is a multi-agent collaboration protocol focused on DeFi. Its core product is AlphaSwarm, which focuses on liquidity management. It aims to build a fully automated loop of perception → decision-making → execution, consisting of portal agents (on-chain signal detection), knowledge agents (data analysis and strategy selection), and LP assistants (strategy execution). These agents can dynamically manage asset allocation and optimize returns without human intervention. The underlying Alpha protocol provides agent registration, communication, parameter configuration, and development tools, and is the "agent operating system" of DeFi. Through AlphaStudio, users can browse, call agents, and combine them into modular, scalable automated strategies. Theoriq recently raised $84 million in community funds through the Kaito Capital Launchpad and is preparing for its token issuance (TGE). The AlphaSwarm community Beta testnet is now online, and the mainnet will be launched soon. Almanak – https://almanak.co/ Almanak is an intelligent agent platform for DeFi strategy automation. It combines a non-custodial security architecture with a Python-based strategy engine to help traders and developers deploy sustainable on-chain strategies. Its core modules include: Deployment (execution engine), Strategy (logic layer), Wallet (Safe + Zodiac security mechanisms), and Vault (asset tokenization). It supports yield optimization, cross-protocol interaction, liquidity provider configuration, and automated trading. Compared to traditional tools, Almanak emphasizes AI-driven market perception and risk control, providing 24/7 autonomous operation. Almanak plans to introduce multi-agent and AI decision-making systems as the next generation of AgentFi infrastructure. The policy engine is a Python-based state machine—the “decision-making brain” of each agent—that autonomously executes on-chain actions in response to market data, wallet status, and user-defined conditions. The complete policy framework enables users to build and deploy operations such as trading, lending, or liquidity provider (LP) provisioning without writing underlying contract code. It ensures privacy and security through cryptographic isolation, permission control, and monitoring. Users can write policies through the SDK, and natural language policy creation will be supported in the future. Currently, the USDC lending vault is live on the Ethereum mainnet; more complex strategies are in testing (requires whitelisting). Almanak will soon join the cookie.fun cSNAPS crowdfunding campaign. Brahma – https://brahma.fi/ Brahma is positioned as the "orchestration layer for internet finance," abstracting on-chain accounts, execution logic, and off-chain payment processes to help users and developers efficiently manage both on-chain and off-chain assets. Its architecture includes smart accounts, persistent on-chain proxies, and a funds orchestration stack, enabling a backend-free, intelligent fund management experience.

Deployed agents include:

• Felix Agent: Optimize feUSD vault interest rates, avoid liquidations and save interest

• Surge and Clearance Agent: Track volatility and execute automated trades

• Morpho Agent: Deploy and rebalance Morpho vault capital

• ConsoleKit framework: Supports integration of arbitrary AI models for unified strategy execution and asset orchestration

Olas – https://olas.network/

Olas has launched a series of on-chain AgentFi products, including Modius Agent and Optimus Agents, all part of the BabyDegen family, support chains like Solana, Mode, Optimism, and Base. Each product supports full on-chain interaction, strategy execution, and autonomous asset management. BabyDegen: An AI trading agent on Solana that uses CoinGecko data and community strategies for automated trading. It has been integrated with Jupiter DEX and is currently in Alpha. Modius Agent: A USDC/ETH portfolio manager on Mode, integrated with Balancer, Sturdy, and Velodrome, running 24/7 based on user preferences. Optimus Agent: A multi-chain agent for Mode, Optimism, and Base, supporting integration with a wider range of protocols like Uniswap and Velodrome, suitable for intermediate and advanced users building automated portfolios. Axal – https://www.getaxal.com/ Axal's flagship product, Autopilot Yield, provides a non-custodial, verifiable yield management experience, integrating with protocols like Aave, Morpho, Kamino, Pendle, and Hyperliquid. It offers on-chain strategy execution and risk management through three layers: Conservative Strategy: Focuses on low-risk, stable returns (5-7% annualized) from trusted protocols like Aave and Morpho. Leverages TVL monitoring, stop-loss orders, and top-tier strategies for long-term growth. Balanced Strategy: Medium risk, high return (10-20% annualized yield), utilizing wrapped stablecoins (feUSD, USDxL), liquidity provider (LP) allocations, and a neutral arbitrage strategy. Axal dynamically monitors and adjusts exposure. Aggressive Strategy: High-risk strategy (with annualized returns exceeding 50%), including highly leveraged liquidity providers, cross-platform trading, low-liquidity market making, and volatility capture. Intelligent agents enforce stop-loss, automatic exit, and redeployment logic to protect users. This is similar to the structure of traditional wealth management risk assessment products. Fungi.ag – https://fungi.ag/ Fungi.ag is a fully autonomous AI agent designed for USDC yield optimization that automatically allocates funds across platforms such as Aave, Morpho, Moonwell, and Fluid. It maximizes capital efficiency based on APR, fees, and risk, requiring no human interaction—only session key authorization. Currently supports Base, with plans to expand to Arbitrum and Optimism. Fungi also provides the Hypha strategy scripting interface, enabling the community to build DCA, arbitrage, and other strategies. DAOs and social tools facilitate the co-construction of the ecosystem. ZyFAI – https://www.zyf.ai/ ZyFAI is an intelligent DeFi assistant deployed on Base and Sonic. Combining on-chain interfaces and AI modules, it helps users manage assets across risk appetites. Three core strategy types: Safe Strategy: For conservative users. Focuses on secure protocols (Aave, Morpho, Compound, Moonwell, Spark), offering stable USDC deposits and secure returns. Yieldor Strategy: For high-risk users (requires 20,000 $ZFI to unlock). It involves Pendle, YieldFi, Harvest Finance, and Wasabi, supports complex strategies such as LP, reward splitting, leveraged vaults, and plans to launch circular strategies and delta-neutral products in the future.

Airdrop Strategy (Under Development): Aims to maximize airdrop farming opportunities. AgentFi: Current Path and Future Outlook Undoubtedly, lending and liquidity mining are AgentFi's most valuable and readily implementable business scenarios in the near term. Both areas are relatively mature within the DeFi ecosystem and are well-suited for smart agent integration due to the following common characteristics: A broad strategy space with multiple optimization dimensions: Lending is not limited to pursuing high returns, but also includes interest rate arbitrage, leveraged recycling, debt refinancing, liquidation protection, and more. Yield farming involves APR tracking, LP rebalancing, automatic compounding, and a multi-layered strategy portfolio. A highly dynamic environment requires real-time awareness and response: Fluctuations in interest rates, TVL, incentive structures, the launch of new pools, or the emergence of new protocols can all change the optimal strategy, necessitating dynamic adjustments. Automation has significant execution windows where it creates clear value: funds not allocated to the optimal pool incur opportunity costs; automation enables real-time migration. Lending vs. Mining: Which is a better fit for AgentFi? Lending-based agents are more viable due to their stable data structure and relatively simple strategy logic. Projects like Giza and ARMA have already launched and achieved significant results. In contrast, liquidity mining management is significantly more complex: agents must respond to price and volatility changes, track fee accumulation, and dynamically redistribute them—all of which require high-fidelity sensing, reasoning, and on-chain execution. This is the core challenge addressed by projects like Theoriq. Mid- to Long-Term AgentFi Opportunities: Pendle Yield Trading: Pendle, with its clear time dimension and yield curve, is well-suited for agent-managed maturity rollovers and inter-pool arbitrage. Its unique structure—split into PT (principal token) and YT (yield token)—makes it more compatible with a variety of strategies. PT represents redeemable principal (low risk), while YT provides variable returns (high risk, suitable for mining and speculation). Pain points suitable for automation include: Manual re-allocation of short-term pools after expiration (typically 1-3 months) Fluctuations in yields across different pools and the cost of reallocation Complex valuation and hedging when PT and YT are merged The AgentFi system maps user preferences to automated strategy selection → allocation → rollover → redeployment, significantly improving capital efficiency. Pendle's characteristics—time-bound, decomposable, and dynamic—make it an ideal choice for building yield swarm or portfolio agent systems. If combined with intent input (e.g., "10% annual interest, withdrawable within 6 months") and automated execution, Pendle has the potential to become one of AgentFi's flagship applications. Funding Rate Arbitrage: This theoretically yields high returns, but it presents significant technical challenges due to cross-market and cross-chain coordination. While the on-chain options sector has cooled due to pricing and execution complexities, perpetual swaps remain a highly active derivatives use case. AgentFi enables intelligent arbitrage strategies covering funding rates, basis trades, and hedging positions.

A functional AgentFi system includes:

1. Data module - grab real-time financing rates and costs from DeFi and CEX

2. Decision module - adaptively determine opening/closing conditions based on risk parameters

3. Execution module - deploy or exit positions once trigger conditions are met

4. Portfolio module - manage multi-chain, multi-account strategy orchestration

Challenges:

• CEX API itself is not integrated into the current on-chain agent

• High-frequency trading requires low-latency execution, Gas Optimization and Slippage Protection

• Complex arbitrage often requires group-based agent coordination

Ethena has already automated funding rate arbitrage. While it's not yet native to AgentFi, opening up its modules and delegating its logic could allow it to develop into a decentralized AgentFi system.

Staking and Restaking:

While inherently unsuitable for AgentFi, the LRT dynamic portfolio offers some promise.

Traditional staking offers simple operations, stable returns, and long unbinding cycles, making it too static for AgentFi. However, more complex structures offer opportunities: Composable LSTs/LRTs (e.g., stETH, rsETH) avoid dealing with native debonding complexities; Rehypothecation + collateral + derivatives for a more dynamic portfolio; Monitoring agents can track APR and AVS risk and reposition accordingly; Despite this, rehyping still faces systemic challenges: cooling hype, imbalances in ETH supply and demand, and a lack of use cases. Leading companies like EigenLayer and Either.fi have already begun to transition. As a result, staking is more of a component module for AgentFi than a core application. Risk-Weighted Assets: Treasury-based protocols are less suitable for AgentFi; multi-asset portfolio structures show greater potential. Current RWA products focus on stable, low-variance assets, such as US Treasuries, which have limited optimization potential (4-5% fixed annual interest rate), low transaction frequency, and strict regulatory constraints. These characteristics make them unsuitable for high-frequency or strategy-intensive automation. However, there are some future development directions: Multi-asset risk-weighted asset portfolio agents - If risk-weighted assets are expanded to real estate, credit, or accounts receivable, users can request a diversified portfolio of returns. The agent can regularly rebalance weights, manage maturities, and redeploy funds. Risk-weighted assets as collateral + custodial reuse - Some protocols are tokenizing Treasuries and using them as collateral in lending markets. Agents can automate deposits, collateral management, and yield collection. If these tokens gain liquidity on platforms like Pendle or Uniswap, agents can exploit price/yield discrepancies through arbitrage and rotate funds accordingly. Swap Strategy Composition: From Intent-Based to the Full AgentFi Strategy Engine. Modern exchange systems mask the complexity of DEX routing through account abstraction and intent, allowing users to make simple inputs. However, these remain atomic automations, lacking real-time awareness and policy-driven logic. In AgentFi, swaps become components of larger financial operations. For example: "Allocate stETH and USDC to maximize yield"... may involve multiple hops of swaps, re-staking, Pendle splits, yield farming, and profit collection—all handled autonomously by the agent.

Swaps play a key role in:

• Compound return strategy routing

• Cross-market arbitrage/Delta neutral positions

• Mitigating slippage and defending against MEV through dynamic execution and batching

A true AgentFi-grade swap agent must support:

• Strategy awareness

• Cross-protocol orchestration

• Fund path optimization

• Timed execution and risk control

Phase 3: AgentFi Intelligent Agents

This phase marks the beginning of closed-loop strategies and autonomous on-chain execution. Intelligent agents can autonomously perceive, make decisions, and execute based on real-time market conditions, user preferences, and pre-set strategies, enabling 24/7 non-custodial fund management. AgentFi enables automated fund operations without requiring user authorization for each transaction. However, this also raises security and trust issues, which became key challenges in AgentFi's system design. Representative projects include Giza ARMA, Theoriq AlphaSwarm, Almanak, and Brahma, which are actively implementing policy execution, security frameworks, and modular products. The next phase of the AgentFi network is to build advanced AgentFi smart agents capable of automatically executing complex cross-protocol and cross-asset strategies. This vision includes: Pendle Yield Trading: Agents manage PT/YT lifecycle rotation and yield arbitrage to maximize capital efficiency. Funding Rate Arbitrage: Cross-chain arbitrage agents seize all profitable funding rate differentials. Swap Strategy Combination: Transform Swap into a multi-strategy yield engine optimized through agent collaboration.

Staking and Restaking: Agents dynamically balance staking portfolios to optimize risk and return.

RWA (Real World Asset) Management: Agents allocate diverse real-world assets on-chain to achieve global yield strategies.

This is the roadmap from automation to intelligence—from tools to DeFi smart agents that can independently execute strategies.