From Random Flows to Intelligent Agents: The AI Economy in 2026 and the Revolution in Scientific Discovery

The transition unfolding in 2026 marks a profound turning point in the history of artificial intelligence. It’s not just about the technical evolution of models but a fundamental reshaping of how AI positions itself within the economic and scientific fabric. While previous years were dominated by the impressive generative capabilities of large neural networks, we are now entering an era where AI agents operate as autonomous entities capable not only of processing but also of acting, negotiating, and discovering.

This paradigm shift reveals a curious phenomenon: the same AI that “hallucinates” and makes reasoning errors is producing tangible scientific discoveries. Organizations like a16z Crypto have identified three pillars supporting this transformation by the end of the year: a new scientific research paradigm based on agent collaboration, a revolutionary identity system (KYA - Know Your Agent), and a completely restructured economic model to address the “invisible tax” on open networks. These three changes do not exist in isolation — each depends on the others to fully materialize.

Scientific Transcendence: When AI Agents Take Responsibility for Discovery

AI-assisted research is evolving drastically this year. We are no longer talking about support tools — literature summarizers, automatic code generators — but systems capable of formulating original hypotheses, conducting complete experiments, and, most importantly, interpreting their own failures. The qualitative leap lies in moving from “linear prompt engineering” to recursive, hierarchical architectures known as AWA (Agents Wrapped by Agents).

Overcoming Pattern Matching Limits

Scott Kominers, a16z researcher, insightfully described this advance: AI models are no longer limited to understanding and executing direct instructions. They can now receive abstract directives — like “guide a PhD student through a comprehensive investigation” — and produce genuinely novel, functionally correct responses. This represents a breakthrough beyond what was once called “random parroting,” revealing a capacity for slow, reflective reasoning similar to the human “systematic mind.”

The previously criticized feature of large models — “hallucination” — is being reinterpreted in scientific discovery as a mechanism of “generative exploration.” When researchers at the University of Washington used this “creative fantasy” capacity to generate over a million protein structures not found in nature, they discovered a new luciferase with catalytic properties comparable to natural enzymes but with superior substrate specificity. Similarly, researchers applying physics-informed neural networks (PINNs) uncovered new unstable singularities in Navier-Stokes equations — the modern equivalent of the classical Stokes law problem in fluid dynamics — revealing patterns previously unknown in complex fluid motion.

The core of this transformation is simple but powerful: allow the model to “dream” in the abstract space to generate high-entropy conjectures, which are then filtered using rigorous logical verification. This turns hallucination into a tool for discovery.

Recursive Architecture: How Agents Collaborate

To harness this emerging capacity, scientific workflows are abandoning linearity in favor of complex hierarchies. The AWA architecture is not just a dialogue among multiple agents — it’s a recursive control system where tasks are decomposed, distributed, and validated in layers.

The “Orchestrator-Executor” pattern is currently the most common implementation. A “Principal Investigator” agent maintains the global context and objectives, decomposing tasks for specialized “Executors.” Data from Anthropic shows a remarkable result: a system with Claude Opus as coordinator and multiple Claude Sontnet agents as executors outperforms a single Claude Opus in complex research tasks — with an improvement of 90.2%. This gain mainly results from “context isolation” — the main agent does not process redundancies of each subtask, maintaining clarity of reasoning.

Another critical feature is the recursive self-improvement cycle. When an agent fails at a task, error data is fed back to a “critical” agent for analysis and correction. The MOSAIC framework (Multi-Agent System for AI-Driven Code Generation) significantly increases scientific code generation accuracy by introducing specialized agents for self-assessment and principle formulation — simulating human thought processes in the face of experimental failure.

The “AI Scientist”: A Complete Discovery System

The most emblematic example of this architecture is Sakana AI’s “AI Scientist,” launched in 2025. This system automatically executes the entire scientific discovery cycle: from generating creative ideas (based on models like NanoGPT and literature research), through experimental iteration (with automatic code correction via tools like Aider), to writing full LaTeX articles with automatic references, and finally peer review simulated according to top conferences like NeurIPS.

The economic result is impressive: the computational cost to generate a complete research paper is around $15. Even more surprisingly, a paper generated by this system, “Compositional Regularization,” successfully passed peer review at an ICLR workshop. Despite still having limitations — such as reference hallucinations and logical errors — this case demonstrates that AI has transcended assistance to autonomous execution of complex scientific processes.

Identity Commands: Building Trust in an Economy of Machines

As agents gain rights to execute transactions and actions in the real world, the digital economy faces an unprecedented identity crisis. Sean Neville, CEO of Catena Labs, warned that the number of “non-human identities” in finance has already dramatically surpassed the number of human employees — in some statistics, up to 96 to 1. These agents — without traditional bank accounts, without real identity verification, but operating at machine speed — represent a regulatory compliance vortex.

The Rise of Non-Human Identities and Their Risks

Forty-five percent of financial institutions internally admit to the existence of “shadow AI agents” that are unapproved, creating “islands of identity” outside formal governance. Tangible risk scenarios include: an resource-optimizing cloud agent autonomously purchasing expensive instances; a trading robot executing incorrect sell instructions during extreme volatility. The key question is responsibility: when an agent violates rules, who is responsible? The developer? The manager? The model provider? Without a clear identity system, these responsibilities remain nebulous.

KYA: A Foundation of Trust

KYA (Know Your Agent) is not just about issuing IDs. It’s a comprehensive digital identity system that includes subjects, credentials, permissions, and reputation history.

KYA’s structure rests on three pillars:

Principal Subject: The entity legally responsible for the agent. The agent must be cryptographically linked to a verified person or company account with KYC/KYB.

Agent Identity: Based on Decentralized Identifiers (DIDs). These are cryptographically generated, immutable, and portable across platforms — creating real portability.

Mandate and Authorization: Verifiable Credentials (VCs) that explicitly declare the agent’s rights. For example: “This agent is authorized to represent Alice in Amazon purchases, with a maximum limit of $500.”

The trust mechanism works as follows: when an agent initiates a transaction, it presents a VC. The verifier does not trust the agent itself — it trusts the digital signature of the VC. If the signature comes from a trusted issuer, the transaction is approved. This creates a clear “chain of trust”: the bank trusts the company → the company issues a VC to the agent → the merchant verifies the VC → transaction authorized.

Protocols Supporting KYA

The technical standards battle is in full swing. The Skyfire protocol launched KYAPay, an open standard where the token consists of two components: identity information and payment capability. This allows an agent to complete a “visitor checkout” without manual form filling.

Catena Labs, founded by USDC architect Sean Neville, developed ACK (Agent Commerce Kit), described as the “HTTP of intelligent agent business.” ACK uses W3C DID standards and allows agents to directly control smart contract wallets, offering superior security over traditional API keys.

Google launched the Agent Payment Protocol (AP2), managing permissions via “Approval Letters” and collaborating with Coinbase on the AP2 x402 extension, which embeds encrypted payment standards directly into the protocol.

Reputation and Risk Control

KYA also serves as the basis for reputation systems. The ERC-7007 standard records each successful interaction — punctual payments, high-quality code — on the blockchain, forming a verifiable curriculum. Financial institutions are implementing smart access portals that, if an agent’s behavior deviates significantly (e.g., abnormal high-frequency transactions), can immediately revoke VCs, triggering a real-time “digital suppression.”

Economic Restructuring: From Attention Model to Value Model

a16z’s Liz identified a fundamental problem: AI agents are imposing an “invisible tax” on open networks. They mass-extract data from content sites, systematically bypassing advertising and subscription models that sustain content creation. If this parasitic relationship isn’t addressed, the content ecosystem will be drained.

The “Great Disconnection”: Data Erosion of Traffic

In 2025, the digital publishing industry experienced a phenomenon called the “Great Disconnection.” Search volumes increased, but clicks to sites plummeted. a16z predicts a 25% reduction in search engine traffic by the end of 2026. Similarweb data shows that the rate of search queries without clicks rose to 65% — users get answers and never visit the original page.

Metrics are even more severe: click-through rates (CTR) dropped sharply when AI summaries appeared above search results. DMG Media reported an 89% decline in clicks, with the traditional first search result losing 34.5% of its previous traffic.

A New Model: Usage-Based Payments

To address this crisis, the industry is moving away from static annual data licenses (like Reddit-OpenAI agreements) toward usage-based compensation. Perplexity AI’s Comet Plus exemplifies this: it established a revenue pool of $42.5 million. When an AI agent cites content from an publisher or accesses pages on behalf of a user, revenue sharing is triggered. Publishers can receive up to 80% of this split — explicitly recognizing the value of “machine access.”

Technical Standards: Native Microtransactions

To expand this model across the open network, a series of technological standards are being implemented. The HTTP status code 402 — historically dormant — has been activated via the x402 protocol, establishing the “Native Machine Payment” standard.

The flow is elegant: agent requests resource → server returns 402 Payment Required with price (e.g., 0.001 USDC) → agent automatically signs payment via blockchain L2 (Base, Solana) or Lightning network → server verifies and releases data. Traditional payment gateways cannot process such tiny amounts, but x402 combined with low fees makes payments by nanointeraction feasible.

The TDMRep (Text Data Mining Protocol) allows sites to declare in robots.txt or HTTP headers: “TDM rights reserved, payment required.” This gives agents a clear binary signal. The C2PA (Content Provenance and Authenticity Alliance) embeds cryptographic “content proofs” verifying origin, ensuring the attribution chain remains intact even when consumed by AI.

Programmable Intellectual Property

An even more ambitious reform is the tokenization of intellectual property via the Story Protocol. Creators register works as “IP assets” on the Story Network, with “programmable IP licenses” embedded. When AI agents use this data, smart contracts automatically execute terms (e.g., “5% royalties for commercial use”) and distribute profits autonomously. This creates a high-liquidity IP market, eliminating legal intervention.

From SEO to AEO: A Paradigm Shift in Marketing

In 2026, marketing focus shifts from SEO to AEO (Agent Exclusive Access, or alternatively “GEO” — Geographic Engine Optimization for agents). The goal is no longer to be the “top search result” — it’s to be cited by the AI agent or become the “preferred data source” in its reasoning process. The emerging advertising model will be “contextual injection”: brands compete to enter the agent’s reasoning chain, so a travel planning agent “remembers” that a particular hotel is the best choice during its analysis.

Conclusion: A Deep Reconstruction

The technological landscape of 2026 makes one thing clear: the friction between human-centered internet infrastructure and machine-centered needs is forcing a profound rebuild of the digital world.

In science, AI has evolved from assistance to full autonomy. The recursive agent architecture enables mass scientific discoveries at negligible costs, transforming “hallucinations” into mechanisms of creativity and solving complex problems — from fluid dynamics (Stokes law) to protein design — in ways previously impossible.

In identity, KYA emerges as the new frontier of financial compliance, assigning billions of AI agents verifiable legal economic identities, allowing them to navigate value networks securely and frictionlessly.

In the economic model, the digital economy is shifting from attention-based advertising to a value-based model. Native payments and programmable IP form the pathways of this new economy, solving the “invisible tax” on open networks and ensuring data producers remain profitable in the post-click era.

We are witnessing the birth of an agent economy — a world where software not only helps us work but is itself a producer, consumer, and trader. This is not a distant future. It is unfolding now.

About Movemaker

Movemaker is the first officially authorized community organization by the Aptos Foundation, jointly initiated by Ankaa and BlockBooster, dedicated to promoting the construction and development of the Aptos ecosystem in the Chinese-speaking region. As the official representative of Aptos in the Chinese community, Movemaker is committed to creating a diverse, open, and prosperous Aptos ecosystem, connecting developers, users, capital, and multiple ecological partners.

Disclaimer: This article is for informational purposes only and reflects the personal opinions of the author, not the position of Movemaker. It does not intend to provide: (i) investment advice or recommendations; (ii) an offer or invitation to buy, sell, or hold digital assets; or (iii) financial, accounting, legal, or tax advice. Holding digital assets, including stablecoins and NFTs, involves high risk, with significant price volatility and potential total loss. Carefully consider, based on your financial situation, whether trading or holding digital assets is suitable for you. For specific questions, consult your legal, tax, or investment advisor. The information provided (including market data and statistics, if applicable) is for general reference only. Reasonable care has been taken in preparing this data, but we do not assume responsibility for any factual errors or omissions.