By Rustem Zhexembin & Hiroki Kotabe
Thanks to all who contributed to this article: Krakus, Jason, Benedict, Yash, Rishin, Jordan
In 2025, DeFi looks nothing like its early days. The numbers alone tell the story: institutional flows topping $10 billion in a single quarter, and more than 3,000 active protocols across dozens of chains. The Total Value Locked (TVL) across all DeFi protocols reached $160 billion in 2025, up 41% year-over-year, not even mentioning that total DEX and Perps volume exceeds trillions.
As DeFi gets bigger, there’s more you can do, but things also get a lot more confusing. Most people can’t keep up with everything that’s happening on-chain. If we want everyone to take advantage of these new opportunities, we need to build tools that make it much easier to make good decisions. That’s where the future is headed.
At the same time, as AI becomes part of daily life, people are developing new routines around automation. This shift is giving rise to agentic finance, where intelligent agents handle both navigation and execution.
Even simple browser-based agents like Comet show how quickly these tools are evolving. When you run a DeFi flow through a browser agent, like in the example shared by Yash, founder of SendAI, you can see the promise of agentic finance.
The vision is simple, instead of digging through dashboards and X threads, you tell an AI what you want, and it handles the rest. Two types of agents are emerging, AI agents as copilots that guide users across all of DeFi, and more technical AI quant agents that automate professional strategies. Both are early, both are flawed, but together they point to a different way of interacting with decentralized finance.
To keep terms clear, we’ll call the first group copilots and the second quant agents as autopilots.
Think of these agents as personal assistants. You ask in plain English, "top trending tokens? where is the best yield" instead of reading charts and and browsing through different protocols. The agent can answer directly and suggest next steps, so it's like having a knowledgeable friend on call 24/7.
Take &milo for instance. Its copilot mode assists you with decisions, rebalancing, and portfolio insights — keeping you in control while doing the heavy lifting.
With plain-language explanations and smart prompts, &milo helps users understand positions and compare opportunities without digging through dashboards. It’s an early look at how copilot agents are evolving from simple chat helpers to full-featured DeFi guides.
To see how these agents perform in practice, we tried one of the latest releases to get a firsthand sense of how they handle real DeFi tasks.
That said, the agent still has limitations. While it successfully identified trending tokens, it was unable to execute a buy order and failed two transactions due to what it flagged as insufficient balance, even though there was enough SOL to cover the fees.
A similar platform called The Hive connects a swarm of DeFi agents for tasks like bridging, yield strategies, or liquidation defense, all coordinated through a simple chat interface. This network of specialized agents work together to execute complex on-chain operations through natural language commands.
We tested the same buy transaction with The Hive. It did surface WEED as a trending token, but when asked to execute a purchase, it returned the wrong contract address.
Milo shows how agents can bundle portfolio tools into one flow, while The Hive experiments with swarms of specialized agents coordinating on-chain.
As agents get more capable, they’re also starting to specialize. Meridian focuses on the opposite end of the spectrum, helping beginners take their first steps in DeFi. Its mobile-first design and clear prompts make simple actions like swapping, staking, or checking yields easy to follow. It performs well on these core tasks and executes quickly, but more importantly, it understands its boundaries. When asked to do something beyond its scope, it explains why, instead of attempting it blindly. That honesty makes it a dependable starting point for anyone new to on-chain activity.
Benedict, the founder of Meridian, explains:
Meridian enables safe research and execution using natural language. We’ve made the agent’s research capabilities available to the public, for free, at meridian.app. Users who make a free account on the Meridian mobile app can make use of the agent’s swap, multi-swap, and portfolio buy capabilities.
Accounts are currently in private beta. Reach out to @bqbrady on Twitter if you’re interested in giving these a try.
Through our playtesting, we found AI agents focused on navigating through DeFi landscape mostly act as teachers and assistants for now that helps with the most simple actions as swaps. However, reliably handling more complex flows, like providing liquidity or managing leveraged positions, remains a work in progress.
As Rishin Sharma, Head of AI at Solana Foundation, notes:
LLMs will often hallucinate on broad tasks and struggle to perform deterministic actions. On the other hand, function or tool calls, like MCP, might be better suited to translating an action plan into execution.
While LLMs can be good at ideation and guidance, they struggle with precise execution. Making agentic finance reliable will require going beyond LLMs to developing specific function calls, clear enactment policies, verifiability, and safe permission systems. In other words, the agentic execution layer is currently underdeveloped - the AI “brain” needs a better “body” to execute its commands.
If copilot-style agents act like tutors, quant agents are more like autopilots. They don’t just build strategies, they run them. These agents watch markets, test trades, and act at machine speed, putting advanced DeFi strategies on autopilot.
A good example of how this infrastructure is forming comes from SendAI. It isn’t a quant agent itself but rather a toolkit that lets others build them. Its “agent kit” for Solana supports over 60 autonomous actions, things like swapping tokens, launching new assets, or managing lending flows, and connects directly with protocols such as Jupiter, Metaplex, and Raydium. In other words, it gives developers the rails to plug decision-making models into on-chain execution.
Yash, summed up their vision clearly:
We at SendAI firmly believe that every AI agent will have a wallet. We’re building the tooling and economic layer for these agents to take any Solana actions. We’re building the platform for these agents to be context-aware and to support long-running, durable, asynchronous executions for complex workflows.
Other teams are focusing on accessibility. Lomen curates strategies and lets users deploy them with a single click, lowering the barrier for people who want quant-level automation without writing code.
For power users who prefer to design their own systems, tools like Unblinked provide AI-powered strategy sandboxes. Think of it as a Cursor for trading: users can sketch out an idea, run it in a safe environment, and refine it before putting capital at risk.
Then there are platforms leaning on multiple agents at once. Almanak combines coding agents with backtesting agents: you describe a strategy in natural language, the AI generates production-grade code, and then tests it against more than 10,000 Monte Carlo simulations. The result is a strategy that’s battle-tested before ever touching real liquidity.
Finally, some teams are chasing real-time edge in live markets. Giza’s ARMA agent actively reallocates funds between lending protocols to maximize stablecoin yields. Instead of leaving capital parked in one pool, ARMA continuously monitors interest rates, liquidity, and gas costs--and moves funds accordingly. Their flagship agent has managed over $17 million and claims returns up to 83% higher than static positions.
Together, these quant agents save time and unlock sophisticated strategies that would be difficult to manage manually. But they also highlight the fragility of automation: agents still stumble when data is stale, when protocols pause unexpectedly, or when shocks ripple through the market. In other words: they can make you faster, but not invincible.
Spend time with today's agents and you'll notice familiar problems. They sometimes suggest actions that no longer exist, like a pool that's already closed. They rely on data that lags behind reality. When a multi-step plan breaks, they don't adapt, they just keep trying the same thing. Permissions are clumsy, either you give full access to your wallet, or you approve each tiny step. And testing is shallow. Simulations rarely capture real-world chaos like sudden liquidity shifts or governance changes.
One of the biggest shortcomings is that agents are operate like black boxes. Users can’t see which inputs were read, how options were weighed, what checks ran against live state, or why a specific transaction was chosen. There’s no signed, reproducible trace to compare what was promised with what was done. Users are forced to babysit automations, and it makes performance hard to judge. Without a way to verify decisions and prove that actions followed declared policies, users can’t tell reliable systems from just good marketing.
Larger capital will require platforms to shift from “trust us” to “prove it” . It is also the hinge into the infrastructure needed to make agent behavior auditable, policy-governed, and trustworthy at scale.
The core issue is that today’s systems don’t have the tools to make agents trustworthy, consistent, and safe at scale. To solve this, we need infrastructure that can check what agents do, confirm the outcome, and follow the same basic rules everywhere. That’s what will make people comfortable trusting them with real money.
However most people don’t need to see how an agent decides, they just need to know its output is correct, verified, and within safe boundaries. Reliability builds trust faster than visibility ever could.
That’s where verifiable reliability comes in. Instead of logging every internal step, agents should run within clear policies and sanity checks: caps on spending, time windows for actions, and checkpoints before major moves. Under the hood, these policies can be enforced through Trusted Execution Environments (TEEs) or similar systems that prove the agent stayed within its constraints without exposing every detail. The result is simple: outputs that can be audited if needed, and confidence that normal users can feel immediately.
This level of verification doesn’t have to be uniform. Everyday use cases can rely on lightweight safeguards and consistent benchmarks; high-stakes or institutional settings can demand stronger attestation and formal proofs. What matters is that each layer of the stack offers measurable reliability appropriate to its risk level.
The next step is agent-readiness. Most protocols still aren’t designed for agents. They need stable, safe surfaces — actions that can be previewed, retried safely, and executed against consistent data. Permissions should be scoped, not absolute, so agents act within clear limits instead of holding the whole wallet. Until these basics exist, even the smartest frameworks will keep tripping over brittle rails.
Once those foundations are in place, users won’t have to babysit automations. Teams will spend less time debugging and more time building. Outputs will be comparable across vendors because they’re backed by benchmarks, not promises.
The fix is straightforward: make agents provable and protocols agent-ready. Put a policy layer between agent and wallet and require attested execution so every run is traceable from request to result, not a black box. Termina’s SVM Engine is being built with this in mind, giving AI agents a true Solana runtime to model decisions and learn from on-chain data. Expose protocol surfaces that are dry-runnable with clear failure codes, safe to retry without causing duplicate actions, consistent on core schemas (positions, fees, health), and scoped by session-style permissions. Ship these and users stop babysitting, teams cut incidents, and institutions finally get the guardrails and proofs they need.
In the next six months, expect copilots to improve first. Better data pipelines will make them more reliable for everyday users. Within a year, stronger testing standards should allow agents to coordinate across multiple protocols, with humans approving only the big steps. Over the longer term, as infrastructure matures, agents may blur into the default interface for DeFi, not as separate tools, but the way people interact with finance itself.
Agentic finance lowers the barrier to entry and brings automation to the experts. But to work at scale, it needs better tooling: live data, safer permissions, robust testing, and transparent results. Smarter AI alone won't solve this. The real progress will come from building the foundations.
DeFi's next milestone isn't just growth. It's trust in automation. And that milestone will be reached only when AI agents stop being demos, and start being dependable.
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.
Share Dialog
Support dialog
920
