The Arbitrage Era of Harness: Rescuing DeFi from the SaaS Edge
TechFlow Selected TechFlow Selected
The Arbitrage Era of Harness: Rescuing DeFi from the SaaS Edge
Reinventing DeFi, the Renaissance of the AI Narrative.
Navigating Web3 tides with focused insightsAuthor: Zuo Ye Web3
Reinventing DeFi: The Renaissance of the AI Narrative
Looking back over 500 years, labor-capital contradictions under capitalism have consistently been marked by capital’s repeated victories.
On the production side, labor’s involvement has gradually shrunk to the level of operating machines. On the consumption side, users’ value lies in generating usage data for platforms.
Together, these forces sustain corporate valuations in capital markets.
Yet human organizational models remain resistant to full quantification: white-collar KPIs/OKRs still follow hierarchical structures; million-dollar annual salaries and piece-rate wages are both variants of Taylorism.
Without clear formulas, capital cannot assign valuations—undermining capital efficiency. Whether algorithmic stablecoins represent DeFi’s “holy grail” remains uncertain, but computable organizations are undeniably the measuring cup for financial leverage.
Large language models (LLMs) have opted for brute-force token-based scaling. The collapse of security-focused SaaS is merely symptomatic; product design is underway, and replacing niche professional capabilities—then scaling them—is the critical bottleneck. Innovation has entered uncharted territory.
This offers profound insights—especially amid the gradual collapse of DAO-based DeFi models and the bankruptcy of tokenomics.
In one sentence: Why are AI’s organizational and token models more efficient than DeFi’s?
How Did This Begin?
Token commoditization and Agent practicality.
For 300% profit, capitalists will sell their own gallows; for job security, workers will write Skills for Agents.
At the capital level, Skills-enhanced Agents hold sacred status equal to profit itself.
Agents embody the distillation of “human capability” into Skills—and beyond that, human organization transforms into an interaction ritual chain centered on Agents.
From Prompt and Context to today’s Harness engineering, all aim to render human organizational patterns uncharted territory—at minimum, reducing human involvement.
Your next colleague won’t be a robot—it could be an “instinctive capability.”
This isn’t fantasy. Scaling Laws governing data performance are gradually failing. Yet data collection and generation no longer matter most. Before AGI arrives, new valuation benchmarks are urgently needed.
Caption: Content Is No Longer Valuable
Source: @ARKInvest
Starting with Claude’s focus on programming as its first step toward AGI, AI has moved beyond chat-box entertainment to penetrate real-world, mature markets—programming, cybersecurity, and recently, design.
This disruptive innovation will either generate new economic growth or plunge the economy into a permanent low-employment paradigm where tokens get hired and humans get laid off—we’re witnessing this unfold.
But token commoditization is already decentralizing capabilities previously monopolized by large enterprises down to SMEs—and empowering “super individuals.” This is no fantasy.
Take China as an example: daily token throughput surged from 100 billion in 2024 → 100 trillion by end-2025 → now 140 trillion per day. Content and data production are nearing zero-cost.
Note: compute scarcity is relative. While large enterprises no longer monopolize “capability,” they still seek to preserve advantages via “compute” monopolies—yet they cannot halt the inevitable broad commoditization of tokens.
Benchmarking base LLM paradigms varies widely—but the evolution of “how AI helps humans” has long gone underappreciated.
To me, Harness represents a spatial paradigm enabling Agents to focus deeply on tasks within defined boundaries—prioritizing depth over breadth, unlike question-answering systems.
Caption: Agent Evolution Timeline
Source: @zuoyeweb3
From the moment Tab key auto-completion first appeared in coding, it was only a matter of time before humans became AI’s input layer.
Trial-and-error costs have dropped exponentially, opening space for novel human collaboration experiments:
- Software: SaaS—human capability now originates not from people, but from emergent Agents
- Hardware: Compute cards + HBM—data centers now serve AI demands directly
- Space: Harness—not physical spaces for human collaboration, but digital spaces for Agent interaction
- Interaction: The demise of Douyin’s “Doubao Phone”; Google embedding GUI Agents at Android’s OS level
AI’s ability to “say things” holds limited commercial value—text generation cost is trivial for humans. But “doing things” will drive token consumption past image/video generation—akin to AWS selling not servers, but compute time.
AI sells not tokens, but “work capability”—this is the root of SaaS industry anxiety. Sadly, DeFi has already become SaaS—not LLMs.
The SaaS-ification of DeFi Protocols
DeFi isn’t outdated—it’s overly precocious.
AI is reinventing software engineering. It’s not just SaaS being displaced—though SaaS is certainly the most emblematic case.
Even Bloomberg Terminal’s core commercial value lies less in technical sophistication than in information authority—a credibility built over decades of industry relationships and non-standardized data networks.
Agents offer an alternative: inferring future outcomes from data—even risky next steps may outperform peers and yield small profits.
Caption: SaaS Collapsing
Source: @zuoyeweb3
You can view this as Agents cleverly exploiting capital’s profit-seeking nature: you can wait for complete Bloomberg Terminal data—or gamble with fragmented, imperfect data for potential gains.
This isn’t novel. IBKR founder Thomas Peterffy pioneered—or rather, assembled—the first physical trading terminal in finance, originating from a decommissioned P101.
If a data utilization method yields higher profits, you gain access to more data—the flywheel starts spinning.
SaaS monopolized the past; AI sells the future.
Unfortunately, we must now pivot to DeFi. Recall Dune/DeFiLlama’s API paywalls—hoarding golden data while begging for scraps—or Arkham Exchange’s eventual shutdown.
Crypto industry data has never held intrinsic value.
Yet crypto is an openly accessible financial system whose data can be repeatedly learned from. Even pre-AI, fork project velocity had already slowed to monthly cycles; PumpFun-style meme clones can compress replication down to seconds.
Here lies a counterintuitive inference: DeFi serves as the financial system’s live testnet. Today’s AI+DeFi experiments will become tomorrow’s financial evolution blueprint.
- E.g., pre-2008 crisis, unsecured LIBOR transactions “triggered” the financial tsunami; SOFR—based on U.S. Treasury transactions—replaced it, yet DeFi’s over-collateralization ensures finality in liquidations.
- E.g., LLM vendors resist selling tokens purely by consumption volume—they insist on tiered marketing, capability customization, and domain-specific adaptation. Tokenomics has twisted “use value” into pretzel-like complexity.
Crypto tokens obsess over use value; AI tokens obsess over economic value.
Viewed this way, DeFi hacks are merely routine stress tests—external entropy that open systems cannot self-repair.
A darkly humorous Catch-22: absent external signaling, crypto defaults to assuming current conditions are safe; once a security crisis hits, it collapses into centralized handling systems.
For instance, during the Drift incident, public blame oddly shifted onto Circle for delayed freezing.
Caption: Code Cannot Solve Security Problems
Source: @zuoyeweb3
Indeed, prior to AI capability leaps, DeFi had already completed its SaaS-ification—charging solely per transaction, unable to port “finance” directly onto-chain.
RWA tokenization suffers liquidity deficits—DeFi lacks effective solutions.
Yet advancing Agent capabilities offer faint, nascent hope for rewriting DeFi’s rules.
- Tokenomics: Deploy usage across channels, allocating by “capital efficiency”;
- Rule-setting: Mythos delivers security finality; AI firewalls battle zero-day threats;
- Human organization: Excellent—DeFi already operates with a few people managing billions.
The Renaissance of Engineering Narratives
Where does security originate? From Turing machines’ determinism. Where does danger arise? From infinite possibility.
YC’s Garry Tan’s “Fat Skill, Thin Harness” resonates deeply—it’s about establishing foundational rules: “freedom grounded in order.”
Turing machines allow infinite combinations; von Neumann architectures always face memory-compute latency; even LLMs cannot produce true randomness.
In a future where data is worthless, only human behavior can imbue monetary flows with value.
Yet human behavior requires time to be fully learned by AI—and internalized into engineered, codified expressions.
Pursuing the infinite with the finite is ultimately futile: LLMs can never fully eliminate hallucinations. Only when they approach the frontier of “neither AI nor human capability can reach this” can market mechanisms price them—and only then might we truly trust smart contracts.
Today’s smart contracts hardly qualify as successful: The DAO fork, Curve’s programming language bugs, even Drift’s multisig—all prove “humans retain ultimate code control.”
Moral scrutiny holds no economic value. DeFi’s collaborative model collapsed from DAOs to foundations and “teams” because real-world needs—contract upgrades, business partnerships—demand pragmatism.
Humans simply cannot write perpetually secure, dynamically upgradable code—remember, it’s fundamentally impossible.
If never upgraded, Curve’s own experience shows even tech dependency stacks fail.
The present determines the past; the past determines the future.
From Simons’ Medallion Fund to Numerai’s AI strategies, AI in finance is nothing new. Another counterintuitive case: trading signals actually accelerate AI evolution.
Caption: AI and DeFi: 10 Years
Source: @zuoyeweb3
AI models remain computer paradigms—state machines processing signals. Without external signals, they lack internal capacity to simulate the external world. Yann LeCun and Fei-Fei Li’s bets on world models reflect precisely this imperative.
Yet from DeFi’s perspective, autonomous AI trading hinges on Agents learning human intent through behavioral observation—highlighting humanity’s irreplaceable role. Even when Agents replace labor, they do so by mimicking and summarizing human behavior.
Indeed, humans cannot intentionally randomize: even minute deliberate actions exhibit statistical regularity. True randomness stems from physiological traits—e.g., “I physiologically prefer Ethena’s market-making strategy and instinctively reject XX’s arbitrage strategy”—revealing fuzzy preferences.
It’s certain that attempts to make blockchain/DeFi AI infrastructure—deAI/deAgent/deOpenclaw—have suffered tragic failures over the past decade.
Directly retrofitting DeFi’s structures using cutting-edge LLMs—e.g., Mythos-tested contracts gaining default security, with any modification triggering real-time detection and elevated risk scoring—offers promise.
Regarding human organization, AI chooses “no humans—only human capabilities.” DeFi is uniquely suited for this, arguably unmatched: after rule design, DeFi focuses solely on enhancing capital efficiency under security constraints—mirroring autonomous driving’s L1/L2/L3/L4 tiers, inevitably progressing from information authorization → limited fund access → full fund control.
If Agents continuously learn engineering-grade trading skills and curator management abilities, they’ll inevitably surpass humans in trading and yield generation. Unfortunately, accumulated DeFi data hasn’t yet undergone systematic AI learning and training—the current crypto AI landscape remains in a fundraising phase.
Yet I’m highly confident actual fund deployment will be the next dominant wave of AI-driven DeFi transformation—inevitable.
So, after security (contracts) and organization (humans) undergo upgrades, what form will tokenomics take?
- PoW-era tokens were compute-consumption vouchers—functionally identical to today’s AI tokens;
- PoS-era tokens are discounted expected-return vouchers—AI tokens are evolving precisely here (replacing humans is AI’s expression of this economic value);
- AI-era crypto tokens exceed our current engineering scope—predictions remain theoretical and inherently speculative.
Consider Sky’s token distribution controlling channel APYs, or Claude pricing model capability via token consumption: future crypto tokens will likely function as “capital return rate” vouchers.
Crucially distinguish: PoS-era tokens like $ETH anchor expected returns in economic assumptions—prior-experience-based inferences. AI’s engineering rigor, however, pushes DeFi parameters infinitely close to reality—making return and risk rates highly credible and continuously verifiable.
Users could even price tokens based on the underlying DeFi protocol’s LLM/Agent usage and Harness optimization scores—buying if bullish, selling if bearish.
Conclusion
Countless unspeakable anxieties—and humanity’s unpredictable future.
DeFi’s future splits into economic and technical dimensions. Tokenomics currently lacks robust solutions—but security glimmers with promise: Claude Mythos can threaten the world; conversely, it can also safeguard money.
AlphaGo definitively solved Go; Claude definitively solved programming—such scenarios will multiply. DeFi’s contracts, human organization, even financial units of account—all possess theoretical optimization space.
At minimum, humans needn’t fear total replacement. In a data-worthless era, behavior retains meaning. Currently, Agent takeover remains confined to “micro-tasks” and “micro-payments”—repetitive details. We must transform such repetition into value. AI drives data and content value toward zero cost; AI tokens and crypto tokens alike see unit economic value (cost) relentlessly decline—this is the inexorable trend.
Indeed, this marks the first time money truly opens its doors to individuals—whether funding AI work or crypto consumption.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
@zuoyeweb3
