Why does DeFi's next milestone look at AI?

Original Title: DeFi's next milestone: What it'll take for agentic finance to work

Original author: @Lemniscap

Original translation: Ismay, BlockBeats

Editor’s note: When the world of DeFi becomes so complex that even professional users find it hard to grasp, how can we return control to ordinary people?

This article comes from a research paper by Lemniscap, systematically outlining the rise and realities of “smart agent finance.” From &milo, Meridian to SendAI, The Hive, these early products demonstrate how AI can become a new interface for on-chain interactions, while also exposing significant gaps in execution reliability, permission security, and verification mechanisms. The author points out that for DeFi to move to the next stage, the key lies not in smarter models, but in more trustworthy underlying structures—ensuring that every action of the agents is verifiable, traceable, and trustworthy.

This is not only a turning point in the evolution of technology but also an experiment in trust reconstruction. As stated in the text: the next milestone of Decentralized Finance is not a larger scale, but trust in automation.

By 2025, DeFi will be completely different from its early form.

The data itself speaks volumes: institutional capital inflows exceeded 10 billion dollars in a single quarter, with the number of active protocols across dozens of chains surpassing 3000. The total locked value of DeFi protocols in the entire network is expected to reach 160 billion dollars by 2025, a year-on-year increase of 41%; the cumulative trading volume of DEX and Perps is even measured in “trillions.”

As the scale of Decentralized Finance (DeFi) grows, the possibilities increase, but so does the complexity. Most people simply cannot keep up with everything happening on-chain. If we want more people to seize these new opportunities, we must build tools that enable users to make the right decisions more easily - and this is precisely the direction of future development.

At the same time, AI has gradually integrated into daily life, and people have begun to develop new habits around automation. This trend has given rise to “Agentic Finance”—the navigation and execution of financial operations handled by intelligent agents.

Even simple browser-based proxies like Comet demonstrate the rapid evolution of such tools. When you perform a DeFi operation through a browser proxy (as shown in the example shared by SendAI founder Yash), you can see the potential of smart agent finance.

This vision is actually quite intuitive: you no longer need to search through various dashboards or long posts on X, just tell the AI what goal you want to achieve, and it can automatically help you complete the subsequent steps.

Currently, two types of intelligent agents are emerging:

One type is Copilots, which guide users in making decisions throughout the Decentralized Finance world; the other type is Quant Agents, which are more focused on professional automated strategy execution, equivalent to “Autopilots.”

Both are still in the early stages and have their flaws, but they point towards a new direction—a completely different, AI-driven Decentralized Finance interaction.

as a “co-pilot” smart agent

You can think of these smart agents as your personal assistants. You no longer need to sift through charts or jump between different protocols; just ask questions in natural language, such as: “What are the hottest tokens right now?” or “Where are the highest yields?”, and the agent will be able to answer directly and provide the next steps—just like a knowledgeable friend who is always there for you.

Taking &milo as an example, its co-pilot mode can assist you in making investment decisions, rebalancing assets, and gaining insights into your portfolio—allowing you to maintain control while saving tedious operations.

With the help of natural language explanations and smart prompts, &milo can assist users in understanding positions and comparing yield opportunities without having to sift through data on various dashboards. It showcases the evolution of the co-pilot agent from a simple chat assistant to a fully functional Decentralized Finance guide.

To observe the performance of these agents in actual operations, we tried out several newly released products and personally experienced their ability to handle real Decentralized Finance tasks.

The results show that these agents still have limitations. For example, they can successfully identify popular tokens but cannot smoothly execute buy operations; there were also two failed transactions, with the system prompting “insufficient balance,” even though there was actually enough SOL in the account to cover the transaction fees.

Similar platforms like The Hive take a different approach – it combines multiple DeFi agents into a “hive” that can collaboratively accomplish complex tasks such as cross-chain operations, yield strategies, and liquidation defenses, all coordinated through a simple chat interface. This network composed of dedicated agents can perform multi-step on-chain operations using natural language commands.

We tested the same buy order with The Hive. The system did recognize the popular token WEED, but it returned an incorrect contract address when executing the purchase.

Overall, Milo demonstrates how to integrate portfolio management tools into a seamless process, while The Hive explores how to enable multiple specialized agents to work together. As the capabilities of intelligent agents improve, more distinct divisions of labor are also emerging.

For example, Meridian focuses on the user group at the other end - helping beginners take their first steps into Decentralized Finance. It adopts a mobile-first design, combined with clear prompts, making basic operations such as swapping coins, staking, or checking yields easier to grasp.

Meridian performs smoothly and executes quickly on these core tasks, and more importantly, it is very clear about its own boundaries. When users ask it to perform operations beyond its scope, it explains the reasons instead of blindly attempting - this “honesty” makes it a reliable starting point for newcomers exploring the on-chain world.

The founder of Meridian, Benedict, explained:

“Meridian allows users to conduct secure research and operations using natural language. We have made the research function of the agent publicly available for free at meridian.app. Users who register for the Meridian mobile app can use the agent's swap, multi-swap, and portfolio purchase functions. Currently, accounts are still in the closed testing phase, and interested users can contact @bqbrady on Twitter to apply for an experience.”

Through our tests, we found that most AI agents focused on DeFi navigation are still more in the role of “teachers” or “assistants”, mainly helping users complete the most basic operations (such as exchanging tokens).

Further improvements are still needed to enable them to reliably handle more complex processes - such as providing liquidity, managing leveraged positions, etc.

As pointed out by Rishin Sharma, head of AI at the Solana Foundation:

“Large Language Models (LLMs) are prone to hallucinations when handling broad tasks and struggle to execute deterministic operations. A function calling mechanism like MCP may be more suitable for transforming 'action plans' into actual execution. While LLMs perform well at the conceptual and guiding levels, they still fall short in precise execution. To make intelligent financial agents truly reliable, we must move beyond LLMs and develop specific function calling mechanisms, clear execution strategies, verifiability, and secure permission systems. In other words, the current execution layer of intelligent agents is still underdeveloped—AI's 'brain' is smart enough, but it still lacks a 'body' that can act robustly.”

as an “autonomous driving” intelligent agent

If the “co-pilot type” of agent is more like a mentor, then the “quantitative type” of agent is more like an autonomous driving system. They can not only build strategies but also truly execute them—monitoring the market in real time, testing trades, and acting automatically at machine speed, allowing complex Decentralized Finance strategies to enter “fully automated operation” mode.

A typical case that is taking shape comes from SendAI. It is not a quantitative agent itself, but a toolkit that allows others to create these agents. Its “Agent Kit” designed for Solana supports over 60 autonomous operations, including token swaps, new asset issuance, lending management, and can interact directly with mainstream protocols such as Jupiter, Metaplex, and Raydium.

In other words, it provides developers with a “track system” that allows them to directly integrate decision models for on-chain execution.

SendAI founder Yash clearly summarized their vision:

“We believe that every AI agent will have its own wallet in the future. SendAI is building the tools and economic layer needed for this system, enabling these agents to perform any operation on Solana. We are creating a platform that allows these agents to have contextual awareness and supports long-running, persistent, and asynchronous execution of complex tasks.”

At the same time, other teams are trying to make this capability more accessible. Lomen is responsible for selecting strategies and allowing users to “deploy with one click,” lowering the barrier to enjoying quantitative automation without the need to write code.

For the “advanced players” who prefer a more customized system, Unblinked offers an AI-driven strategy experimentation environment. It's like Cursor in the trading field: users can first outline their strategic ideas, run and optimize them in a secure sandbox environment, and then decide whether to invest real money.

Some platforms also choose to call multiple agents to collaborate and complete tasks.

For example, Almanak combines “programming agents” with “backtesting agents”: users describe strategies in natural language, and AI automatically generates production-level code, performing over 10,000 Monte Carlo simulations for backtesting, ultimately producing a “ready-to-go” strategy outcome.

Finally, the team is focusing on real-time market advantages.

The ARMA agent of Giza actively allocates funds between various lending protocols to maximize stablecoin yields. Instead of letting funds stagnate in a single pool, ARMA continuously monitors interest rates, liquidity, and gas costs, dynamically moving assets. Its flagship agent has managed over $17 million in funds, claiming a yield 83% higher than static holdings.

Overall, these quantitative agents significantly reduce time costs, allowing ordinary users to access complex strategies that were originally reserved for professional quantitative teams. However, at the same time, they also reveal the vulnerabilities of automation: when there are data delays, protocol pauses, or severe market fluctuations, the agents may still “stumble.”

In other words, they can indeed make you faster, but they are far from being 'invincible'.

Their dilemma lies in

After interacting with the current smart agents for a period of time, you will notice some similar issues: they sometimes suggest executing operations that no longer exist, such as a liquidity pool that has long been closed; the data they rely on is often lagging behind the real on-chain state; if there is an error midway through a multi-step plan, they do not self-adjust but instead repeatedly attempt the same action.

Permission management is also quite clumsy - either users must grant full access to the entire wallet, or they have to manually approve every minor operation. The testing phase is equally superficial, as the simulated environment struggles to accurately replicate the “real-world chaos” of sudden liquidity changes or governance parameter adjustments on-chain.

One of the most serious problems is that these agents operate almost like a “black box”.

Users cannot know what inputs it has read, how it weighs options, whether it has checked real-time status, and they do not know why a specific transaction was chosen for execution. Without signed verification of operation records, it is impossible to verify the consistency between the “promised results” and the “actual execution.”

Users can only “supervise” the automation process while using it - not only is this inefficient, but it also makes performance difficult to assess.

Without a mechanism that can verify decisions and prove that actions genuinely adhere to established strategies, users will never be able to distinguish between a “reliable system” and “well-packaged marketing.”

For larger-scale capital, DeFi platforms must shift from “trust us” to “please verify.” This is also a key turning point in establishing a “verifiable, governable, and trustworthy” smart agent financial infrastructure.

infrastructure gap

The core issue is that the current system lacks the foundational tools that allow agents to maintain trust, consistency, and security in large-scale scenarios. To address this, we need infrastructure that can verify agent behavior, confirm execution results, and adhere to unified rules across all environments. Only then will people feel comfortable entrusting real money to them.

However, most users do not actually care about the “thought process” of the agent; they just want to confirm that the output is correct, validated, and within safe boundaries. In establishing trust, “verifiable reliability” is more important than “visibility.”

This is precisely the meaning of “Verifiable Reliability”. The agent does not need to record every internal operation, but should operate under clear strategies and reasonable checks: setting spending limits, execution time windows, confirmation nodes before key operations, etc.

At the core, these rules can be ensured through Trusted Execution Environments (TEEs) or similar systems—without exposing all the details, while still proving that the agent indeed adhered to the boundaries. The result is: outputs that can be audited when needed, and operations that ordinary users can trust instantly.

This verification layer does not have to be a “one-size-fits-all”. Lightweight security protection and standardized metrics can be used in everyday scenarios; while high-risk or institutional-level scenarios can require stronger proof and formal verification. The key is that each layer of infrastructure should provide measurable reliability that matches its risk level.

Prepare the protocol for the agent

The next step to be added is to make the protocol “friendly to agents.”

Currently, most DeFi protocols are not designed for smart agents. They need to provide more stable and secure execution interfaces: the ability to preview operations, perform secure retries, and execute based on a consistent data structure. Permission design should also be “scope-limited” rather than “fully open”, allowing agents to operate within defined boundaries rather than having control over the entire wallet.

In the absence of these foundational elements, even the most intelligent agent frameworks can be tripped up by a fragile underlying structure. Once these foundations are in place, users will no longer need to manually monitor automated processes; development teams can reduce debugging time and focus on innovation; and the execution results of different service providers can also be comparable due to shared benchmarks - no longer just a promotional slogan.

must change part

The solution is actually not complicated: make the agents provable and prepare the protocols for the agents. Add a strategy layer between the agents and the wallets, and require that all execution processes are traceable and verifiable, rather than being a “black box operation.”

For example, Termina's SVM engine is built on this concept—it provides a true Solana runtime environment for AI agents, allowing them to model, make decisions, and learn based on on-chain data. At the same time, the protocol party should open up operational interfaces that can be “dry-run,” clear error codes, safe retry mechanisms, consistency of core data structures (positions, fees, health), and session-based permission control.

When these functions are implemented, users will be able to free themselves from the burden of “custodial” proxies; teams can reduce system failures; and institutional investors will finally be able to obtain the security safeguards and verifiable proof they need.

real-time schedule

In the next six months, the fastest improvements are expected to be in “co-pilot” agents. More refined data pipelines will enhance their reliability in everyday use cases.

Within a year, as testing standards improve, agents will be able to coordinate execution across protocols, with humans only needing to approve key steps. In the longer term, as infrastructure matures, intelligent agents may gradually blur into the default interaction layer of Decentralized Finance—no longer separate “tools,” but becoming the primary way people interact with the financial system on a daily basis.

Conclusion

Agentic Finance is lowering the barriers to participation, making automation no longer just a tool for experts. However, to truly operate on a large scale, it needs a better “foundation”: real-time data, more secure permission mechanisms, stronger testing systems, and more transparent execution results.

Relying solely on smarter AI cannot solve these problems. True progress will come from the improvement of the underlying structure.

The next milestone of DeFi is not just the growth in scale, but the trust in automation. And this day will only truly arrive when AI agents are no longer just “concept demonstrations” for show, but become truly reliable executors.