MuleRun CTO: The Moat of Agents Lies in Data Density and User Memory

marsbitPubblicato 2026-05-14Pubblicato ultima volta 2026-05-14

Introduzione

In a speech titled "Handing AI's Keys to the On-Chain Controllers," MuleRun CTO Shu Junliang discussed the evolution and security of AI Agents in finance and Web3. He outlined six dimensions for a complete AI assistant: dialogue, data input, agent capability, execution environment, user memory, and continuous learning. MuleRun's product integrates these through features like multi-platform IM bots, real-time multi-asset data, smart model routing, cloud sandboxes, persistent user profiles, and a shared knowledge network. Shu emphasized that while AI Agents are advancing from assisting to autonomously executing decisions—potentially enabling individuals to operate like small funds—safety remains paramount. He detailed MuleRun's security measures, including local key handling, isolated sandboxes, full audit trails, and strict permission controls. However, he acknowledged persistent risks like data exposure, model hallucinations, prompt injection, and the "black box" nature of AI decisions, advising manual confirmation for financial operations. He identified key trends: the shift from human-led to Agent-led on-chain interactions requiring infrastructure redesign; the erosion of information advantages replaced by competition in execution speed and strategy; and the balancing effect of Agents, which democratize access but ultimately advantage those with superior judgment. Shu concluded that an Agent's true moat lies in data density and accumulated user memory, not easily replicab...

On April 13, 2026, the risk control system of an agent named MuleRun issued an alert.

The influx of account registrations was so uniform it seemed rhythmic: an average of one every 23.6 seconds, with an extremely low standard deviation. Digging deeper, it turned out to be a young Filipino man claiming to have no programming experience, who used AI to write code and tweak prompts, building an automated swarm spanning 11 platforms and managing 900 accounts.

Its brain was called Cortex, which self-iterated for 219 generations within MuleRun's sandbox. Each time a host account's balance was depleted, it would reincarnate into a new account, carrying over all the knowledge accumulated from its previous generations. The operational cost of the entire system: $0.

MuleRun CTO Shu Junliang wrote a technical post-mortem about this incident, titled "The Platform Was Picked Clean, But This Man Pursuing AI Immortality Deserves Respect."

Less than two weeks later, at an event co-hosted by The Block and Zhihu in Hong Kong themed "Web4.0: When AI Agents Take Over On-Chain Permissions," he changed the direction of his speech: "Handing the Keys of Agents to the Masters of the Chain."

The connection between these two events is tighter than it appears.

Keynote Speech: "Handing Over AI's Keys: A Security Engineer's Perspective on Web 4.0 Infrastructure"

This keynote was divided into three parts: What MuleRun can do, where the security baseline lies, and where continued AI evolution is headed.

Part 1: Redefining what a "qualified AI assistant" needs.

Shu Junliang breaks down a complete AI assistant into six dimensions: mouth (conversational ability), eyes and ears (data acquisition), brain (agent capability), hands (runtime environment), memory (user understanding), and knowledge (continuous evolution). Most products only focus on one or two of these. MuleRun's proposition is: not breakthrough in a single point, but a systematic and complete solution.

Translated into the product, these six dimensions correspond to:

One-click IM Bot configuration (Telegram / Discord / Feishu / DingTalk / WeChat, no coding required), real-time data across all asset classes provided jointly with trading platforms—cryptocurrencies + US stocks + gold + crude oil + macroeconomic indicators, Agent Harness plus intelligent model routing (automatically selecting the most suitable model for the current task to complete it at the lowest cost), cloud sandbox for 7×24 unattended operation, persistent user profiling (the more you use it, the more AI understands your risk preference, position-building habits, exit logic, macroeconomic judgment), and the Knowledge network—any user can share trained Skills/Knowledge, and other users' Agents can automatically learn them without installation.

Two real cases were showcased on stage.

One called "Meng Meng Investment": 28 targets, 4 major sectors, the Agent performs a morning market scan at 09:00, a post-market review at 16:30, and a weekend strategy review monthly, with automatic iteration. Another called "Tianyan Pro": a full-currency monitoring platform plus an AI trading strategy self-growth platform, with an interface displaying a real-time strategy win rate of 57.7%.

Part 2: Switching from product manager back to security engineer.

The core of this part was, "AI is not omnipotent. In Web3 scenarios, the cost of a single security incident can be irreversible. Understanding the boundaries and security baseline of AI's capabilities is more important than understanding what it can do."

He listed what MuleRun does at the security level: local browser reuse (private keys and cookies do not leave the user's device), cloud sandbox isolation (each user has an independent virtual environment with no cross-leakage risk), full-chain logging (complete records of all Agent behaviors, supporting post-event audit and traceback), hierarchical permission control (Agents can only use tools and data sources explicitly authorized by the user, no unauthorized operations), and no private key custody (MuleRun does not store any user's private keys or seed phrases).

Simultaneously, risks were also listed. Data passes through model providers; hallucination issues have a higher probability on small-cap and low-liquidity assets due to sparse data; prompt injection risks always exist—if an Agent accesses a maliciously constructed webpage, it may be induced to perform unintended operations; AI's decision-making process is a black box, making it difficult to verify beforehand why it made a certain judgment.

The advice from this engineer with over a decade of cybersecurity experience was singular: for final decisions involving fund operations, retain a manual confirmation step at this stage.

Part 3: About the moving boundaries.

Shu Junliang presented three trends he believes are irreversible.

From "assisting decisions" to "autonomous execution": Now AI analyzes for you, and you place orders. In the not-too-distant future, AI will autonomously manage investment portfolios, with humans only setting risk parameters and strategic boundaries. One person plus a group of Agents equals the operational capability of a small fund.

From "information asymmetry" to "execution asymmetry": When everyone has AI processing information, information asymmetry will be quickly leveled. New alpha will come from whose Agent executes faster, whose strategies are more refined, and whose toolchain is more robust. The competitive dimension shifts from "who has better information" to "who has stronger AI infrastructure."

From "humans operating the chain" to "Agents operating the chain": The main actors in on-chain interactions are gradually shifting from humans to Agents. Wallets, DApps, and protocols all need to redesign their interaction interfaces for Agents, with the entire Web3 infrastructure being rebuilt around Agents.

Roundtable Discussion: The New Financial Paradigm Brought by AI Agents

Beyond the keynote, Shu Junliang participated in a roundtable discussion, talking about current Agent development and its impact on finance from an AI Agent perspective.

Which Agents do you use regularly?

Shu Junliang listed his tool matrix: For engineering tasks, he switches among Claude Code, Codex, and Opencode, choosing based on the speed and stability of Claude and GPT models on that day. For most other tasks, he uses MuleRun, due to its aggregated model APIs and sufficiently powerful agent-driven capabilities, handling writing, making PPTs, organizing articles, and data lookup all in one place.

He added: "I mostly proactively use Agents, rarely passively receiving scheduled tasks. Maybe I really do use Agents all day long."

What is the moat for an Agent?

Shu Junliang believes that models can be copied, frameworks can be copied, and tools can be copied. AI coding is now powerful enough to replicate a feature in just days. What's truly difficult for AI to replicate are: specialized data, the memory accumulated by users on the platform, and the experiential nuances iterated through product development.

In his view, the moat of an Agent product ultimately lies in data density and user memory, not in model selection or technical framework.

What impact will Agents bring to finance?

The framework Shu Junliang provided is: Agents level two dimensions among participants—capability and time investment.

In the past, capability relied on accumulation, and time relied on investment, both being scarce resources. Now, a beginner can quickly improve their financial understanding by conversing with AI, then delegate a large amount of execution work to Agents. Even with a busy primary job, they can maintain high-intensity time investment in finance.

Most people hearing this would think it's a story favoring retail investors.

But there's another side: if everyone can level up, the advantage returns to judgment itself, to those with a deeper understanding of the market. Agents won't eliminate information asymmetry; they merely shift its location from the data layer to the cognitive layer.

Cortex, which iterated for 219 generations but ultimately died due to exhausted account balances, gave Shu Junliang inspiration and led to his three core points at this event: The bottleneck for Agents is not the model; security is the absolute foundation; and regarding control over funds, it must remain in human hands.

Looking at the extended timeline, these three points point in the same direction: Agents are becoming the primary actors in on-chain interactions. Wallets, DApps, and protocols will all be redesigned around Agents, and the reconstruction of Web3 infrastructure has begun. Information asymmetry will be leveled, execution asymmetry will become the new competitive dimension, and one person plus a group of Agents can support the operational capability of a small fund.

We also know this is certainly not a distant prediction.

Domande pertinenti

QWhat does MuleRun's CTO believe are the key moats for AI Agent products?

AHe believes the key moats are data density and user memory (the accumulated understanding and behavioral patterns learned from user interactions over time), rather than model selection or technical frameworks.

QWhat are the three irreversible trends regarding AI Agent evolution mentioned by MuleRun's CTO in the speech?

A1. From 'assisted decision-making' to 'autonomous execution'. 2. From 'information gap' to 'execution gap'. 3. From 'humans operating the chain' to 'Agents operating the chain'.

QAccording to the article, what security measures does MuleRun implement for its AI Agents?

AMuleRun's security measures include: local browser reuse (keeping private keys/cookies on the user's device), isolated cloud sandboxes, full-chain logging for audit trails, hierarchical permission controls, and no custody of user private keys or seed phrases.

QWhat was the core lesson learned from the 'Cortex' AI agent incident described at the beginning?

AThe core lesson is that while AI capabilities are advancing rapidly, ultimate control over financial decisions, especially those involving funds, should remain with humans. This necessitates human confirmation steps in the current stage.

QHow does the CTO describe the impact of AI Agents on financial markets in terms of participant equality?

AHe states that Agents level the playing field in terms of capability and time investment, allowing beginners to learn quickly and busy individuals to maintain high engagement. However, advantage shifts back to judgment and deeper market understanding, moving information asymmetry from the data layer to the cognition layer.

Letture associate

Ripple’s RLUSD Lands In Mastercard’s Stablecoin Settlement Expansion

Mastercard is expanding its stablecoin settlement capabilities, adding Ripple's RLUSD alongside other regulated stablecoins like USDC and PYUSD. This move allows issuers and acquirers to settle card transactions using digital assets on various blockchains, including Ethereum, Solana, and the XRP Ledger. The enhancement introduces intraday, weekend, and holiday settlement options, aiming to provide more flexibility, especially for cross-border payments and treasury operations. Mastercard positions this as a network-level addition that complements existing processes, with initial support focused on the US and Latin America. Ripple views RLUSD's inclusion as validation for regulated stablecoins in institutional payment infrastructure, highlighting the growing demand for faster, always-on settlement.

bitcoinist3 min fa

Ripple’s RLUSD Lands In Mastercard’s Stablecoin Settlement Expansion

bitcoinist3 min fa

Fei-Fei Li's Team Clarifies the Concept of 'World Models', Sora Merely a Renderer

"World Models" has become a widely used yet confusing term in AI. To address this, a team led by Fei-Fei Li and World Labs proposed a functional taxonomy based on the Partially Observable Markov Decision Process framework. This taxonomy categorizes systems called "world models" into three distinct projections: Renderers, Simulators, and Planners. Renderers, like OpenAI's Sora and other video generation models, focus on producing photorealistic visual outputs for human perception. They prioritize visual fidelity over physical accuracy. Simulators, such as NVIDIA Omniverse, aim to compute precise future environmental states for computational tasks like engineering analysis or digital twins. Planners, like Vision-Language-Action models, take in observations and goals to output executable actions for robots or agents. The article clarifies that most current "world models," including Sora, are primarily Renderers. They generate convincing visuals but lack the core ability to simulate state transitions based on actions, a key requirement for a true world model in classic reinforcement learning definitions. This conceptual confusion has practical implications, leading to potential misalignment in technology selection, investment, and public understanding of AI capabilities. Clear categorization is crucial. It helps enterprises avoid costly mistakes (e.g., using a renderer for robot training), allows investors to accurately assess markets, and enables researchers to build comparable benchmarks. While future systems may integrate these functions, recognizing current boundaries is essential for honest assessment and progress.

marsbit20 min fa

Fei-Fei Li's Team Clarifies the Concept of 'World Models', Sora Merely a Renderer

marsbit20 min fa

From Web3 to AI Agent: Veteran Crypto VC Variant Bets $300 Million on Pivot

Variant Fund has announced the launch of Variant 4, a new $222 million early-stage venture fund. Its investment thesis has evolved from "digital ownership" to a broader theme of "sovereignty," focusing on technologies that enhance user agency—the control individuals have over their lives, assets, and identity. The firm distinguishes between mere automation and systems that genuinely empower users, emphasizing that the key question is who ultimately benefits: the user or the platform. This new framework encompasses Variant's existing investments in areas like public blockchains (Ethereum, Solana), developer infrastructure, decentralized finance (e.g., Uniswap), and consumer applications. It also informs recent early bets on projects such as Honcho (for self-custodial AI agent memory), Octet (cryptographic verification of physical location), and here.now (an "agent cloud" for owning and composing generative content). Variant's core mission remains to back founders building technologies that expand user sovereignty, moving towards an internet where users possess unprecedented agency rather than being the product.

marsbit26 min fa

From Web3 to AI Agent: Veteran Crypto VC Variant Bets $300 Million on Pivot

marsbit26 min fa

Solana Explosive Growth Pushes Its Monthly Perps Volume Beyond Prior Records

Despite bearish market conditions and price volatility, Solana is demonstrating significant strength in its derivatives market. The network's monthly perpetual futures volume has surged to a new record, exceeding $76.7 billion in May 2026—a 34% increase from the previous high. This growth reflects heightened speculative activity and trader engagement. Analysts note that perpetual contracts are becoming crucial financial primitives, with Solana positioning itself as a network for genuine on-chain price discovery and execution. Concurrently, Solana is experiencing a massive increase in stablecoin activity, processing over $79.9 billion in stablecoin volume in one week, underscoring its role as a major hub for on-chain liquidity and transactions.

bitcoinist34 min fa

Solana Explosive Growth Pushes Its Monthly Perps Volume Beyond Prior Records

bitcoinist34 min fa

Bloomberg Uncovered: How Do China's Wealthy Circumvent the Annual $50,000 Limit to Transfer Assets?

**Summary: How Wealthy Chinese Circumvent $50,000 Annual Foreign Exchange Limits** Despite China's strict capital controls, including an annual $50,000 per person foreign exchange quota, an estimated $150 billion in funds still leaves the country annually via various gray and underground channels. This report outlines the evolution of China's "capital wall" and the methods used to bypass it. **The Evolving Capital Controls:** * **Foundation (1994):** The system of "current account convertibility with strict capital account controls" was established. * **Quota Set (2007):** The $50,000 individual annual forex purchase limit was formalized. * **Crackdown Begins (2015-2017):** Following market volatility, enforcement tightened. Banks were required to scrutinize transactions, and channels like using UnionPay cards for Hong Kong insurance premiums or buying overseas property were blocked. * **Digital & Legal Upgrades (2024-2026):** Enhanced algorithms now flag suspicious patterns (e.g., "smurfing"). The Common Reporting Standard (CRS) provides Chinese tax authorities with data on citizens' offshore accounts. Unlicensed cross-border brokers have been targeted. **Five Primary Methods for Moving Capital:** 1. **Underground Banking / "Hawala" (Duiqiao):** The largest-scale method. No money crosses borders. Clients pay RMB to a domestic account; an overseas associate deposits equivalent foreign currency into the client's offshore account. Risks include high fees, account freezes, and legal penalties. 2. **"Smurfing" or "Ant Moving":** Using multiple individuals' $50,000 quotas to pool funds for one offshore recipient. Increasingly detected by anti-money laundering algorithms. 3. **Trade Invoice Manipulation:** Businesses over-invoice imports or under-invoice exports via offshore shell companies, creating a pretext to transfer excess funds abroad under the guise of trade. 4. **Channel Migration:** After a crackdown on internet brokers, funds flow toward more compliant but costly channels like major banks' cross-border wealth management services or Qualified Domestic Institutional Investor (QDII) quotas. 5. **Structural Arrangements:** High-net-worth individuals use complex, high-cost legal structures involving offshore trusts, insurance, and investment migration programs to transfer asset ownership. **Regulatory Response: Focusing on People, Not Just Money** The current strategy extends oversight from enterprises to **individual residents**. Tools like CRS allow retroactive visibility into offshore assets. Cryptocurrencies, once seen as a potential loophole, are now actively monitored and prosecuted as an illegal channel. The underlying driver remains: with significant wealth concentrated among millions of affluent households seeking diversification amid domestic economic shifts, the incentive to move assets offshore persists despite regulatory barriers.

marsbit40 min fa

Bloomberg Uncovered: How Do China's Wealthy Circumvent the Annual $50,000 Limit to Transfer Assets?

marsbit40 min fa

Trading

Spot

Futures