On the eve of the AI Agent economy: Structural changes in 2026

Author: @BlazingKevin_, The Paintings at Movemaker

Original title: 2026 AI Agent Economic Outlook: Reshaping AI Identity and Network Value Flow

Introduction: Structural transition from generative AI to “agent behavior”

In 2026, the field of artificial intelligence will experience a structural shift from “generative ability” to “agent mobility.” If 2023-2024 is about the amazing language generation capabilities of big language models, then 2026 will mark the formal establishment of the “AI Agent Economy.”

Based on the predictions and analysis of the a16z Crypto research team, our further research revealed that 2026 will be a year of deep integration of AI as a productivity tool with Crypto as a value allocation layer.

AI is no longer just a passive tool that responds to human instructions, but an active participant with the ability to reason, plan, trade, and make independent discoveries.

Based on a16z Crypto's outlook report, the three core trends reshaping the AI+Crypto landscape in 2026 are:

1. A new paradigm for scientific research: moving from a single agent to an “Agent-Wrapping-Agent”.

2. The financial infrastructure revolution: from KYC to KYA (Know Your Agent).

3. Economic model restructuring: Addressing the “hidden tax” crisis faced by open networks through payment and programmable IP.

These three trends are not isolated: the shift in research paradigms relies on advanced collaboration between agents; advanced collaboration requires agents to have verifiable identities (KYA); and identity-owned agents must follow new value exchange agreements when obtaining data.

1. The Neo-Erudicist Era: “Agent-Wrapping-Agent” Architectures in Advanced Research

Starting this year, the definition of “AI-assisted research” will take a qualitative leap forward.

We're not talking about simple literature searches or text abstracts, but rather witness AI systems that can do substantive reasoning, generate hypotheses, and even solve doctoral-level problems on their own.

The core impetus for this transformation is a shift from single-model linear hint engineering to complex, recursive AWA workflows.

1.1 Breakthrough in reasoning ability: crossing the boundaries of pattern matching

Scott Kominers of a16z points out that AI models are evolving from simply understanding instructions to being able to take abstract instructions (like mentoring PhD students) and return novel and correctly executed answers. Recent technological advancements show that AI models are breaking through the “random parrot” ceiling, showing slow, thoughtful reasoning abilities similar to human “systems” thinking.

1.1.1 “The illusion of usefulness”

As reasoning skills increase, a new “erudite” research style is taking shape. Scott describes this style as “using AI to cross disciplinary boundaries and speculate on the deep connections that may exist between topology and economics, biology, and materials science.”

The “illusory” nature of the big model, which has been criticized, is being restructured into a “generative exploration” mechanism in the context of scientific discovery:

• Protein design case: University of Washington researchers used the “Whole-Family Illusion” (concept) to generate more than 1 million unique protein structures that don't exist in nature. Among them, the novel luciferase screened out is comparable to natural enzymes in terms of catalytic activity, but has higher substrate specificity.

• Hydrodynamic discoveries: Through physical information neural networks (PinNs), researchers have discovered new unstable singularities in the Navier-Stokes equations (Navier-Stokes equations), which reveal previously unknown patterns in fluid motion.

The core of this research style is to allow models to “fool around” in an abstract space to generate high-entropy conjectures, and then use strict logical validators to screen conjectures.

1.2 Detailed explanation of AWA architecture

In order to harness this powerful ability to reason and generate, research workflows are shifting from flat to hierarchical. AWA refers not only to conversations between multiple agents, but also to a recursive, hierarchical control structure.

1.2.1 The orchestrator - performer mode

This is currently the most mainstream AWA implementation model. A “lead researcher” agent is responsible for maintaining the global context and research goals, breaking down and distributing tasks to a dedicated set of “performer” agents.

• Architectural advantage: According to Anthropic data, multi-agent systems composed of Claude Opus as the leading agent and Claude Sonnet as sub-agent performed 90.2% better on complex research tasks than a single Claude Opus Agent.

  This performance improvement is mainly due to contextual isolation — the leading agent doesn't need to process redundant information for each subtask, thereby maintaining clarity of reasoning.

1.2.2 Recursive Self-Improvement and the MOSAIC Framework

Another key feature of the AWA architecture is the introduction of the Reflexion (reflection) cycle. When the underlying agent fails to perform a task, the error information is fed back to a “critic” agent for analysis and correction.

The MOSAIC framework (Multi-Agent System for AI-Driven Code Generation) significantly improves the accuracy of scientific code generation without relying on verification test cases by introducing dedicated “self-reflection agents” and “principle generation agents”. This closed loop of “trial and error - reflection - retry” simulates the thought process of human scientists when faced with the failure of an experiment.

1.3 Case Study: Sakana AI's “AI Scientist”

The most notable AWA use case in 2025 was “The AI Scientist” system released by Sakana AI. It is a system designed to fully automate the entire life cycle of scientific discovery.

1.3.1 Fully automated scientific research closed-loop process

1. Idea generation: The system is based on a starter code template (such as nanoGPT), uses LLM as a “mutation operator” to brainstorm diverse research directions, and calls the Semantic Scholar API to search literature to ensure novelty.

2. Experimental iteration: An “experimenter” agent writes and executes code. If the experiment fails, the system captures the error log using the Aider tool and fixes the code on its own until a visual chart is obtained.

3. Paper writing: The “writer” agent writes a complete scientific paper using LaTeX, covering abstracts, methods, experimental results, and independently searches for references to generate BibTeX.

4. Automated peer review: The generated paper is submitted to a simulated “reviewer” agent that scores according to the standards of top conferences such as NeurIPS. The system can even be revised in multiple rounds based on judges' opinions.

1.3.2 Economic efficiency and quality

The economic efficiency of the “AI Scientist” system is astonishing: the computational cost to generate a complete research paper is only about $15. The system-generated paper “Compositional Regularization” was even successfully peer-reviewed at the ICLR seminar. Although there are still limitations such as citation illusions and logic flaws, this case shows that AI already has the ability to not only assist in research, but also perform the complete research process.

2. The Order of Identity: From KYC to KYA

As agents are empowered to perform tasks and transactions, the digital economy is facing an unprecedented identity crisis. Sean Neville (CEO of Catena Labs) warns that the number of “non-human identities” in the financial services sector has reached 96 times the number of human employees, and even as high as 100:1 in some statistics. These agents — without bank accounts, without real name authentication, and running at machine speed — are huge compliance black holes. The industry is urgently shifting from traditional KYC to KYA (Know Your Agent).

2.1 Outbreak and risk of non-human identities (NHI)

2.1.1 “Shadow AI” and the 96:1 imbalance

45% of financial services institutions acknowledge the existence of unapproved internal “shadow AI agents.” These agents create “identity silos” outside of formal governance frameworks.

• Risk scenario: A testing agent for cloud resource optimization may independently purchase expensive reserved instances without human intervention; or a trading robot triggers false selling instructions when the market fluctuates.

• The attribution conundrum: Who is responsible when agents break the rules? Was it the engineer who developed it? The manager who deployed it? Or is it the vendor that provides the basic model? Without KYA, these responsibilities can't be defined.

2.2 The KYA Framework: The Foundation of Trust for the Machine Economy

KYA does more than just issue ID cards; it establishes a complete digital identity system that includes subjects, credentials, authority, and reputation.

2.2.1 The Three Pillars of KYA

1. Subject: The entity that is legally responsible for the Agent. The agent must be cryptographically linked to a KYC/KYB-verified human or business account.

2. Agent identity: A unique digital identity based on a decentralized identifier. DIDs are cryptographically generated, immutable, and portable across platforms.

3. Mandate/Delegate (Mandate/Delegate): A permission statement issued through verifiable credentials (VCs). For example, a VC could state: “The agent is authorized to spend on behalf of Alice on Amazon up to a limit of $500.”

2.2.2 Cryptographic binding and trust chains

When an agent initiates a transaction, it presents a VC. The verifier does not need to trust the agent itself; it only needs to verify whether the digital signature on the VC comes from a trusted issuer. This mechanism creates a “chain of trust”: the bank trusts the enterprise -> the enterprise issues VC to the agent -> the merchant verifies the VC -> the transaction passes.

2.3 Protocol Stack Dispute: Standard-setting for Agent Identity

2.3.1 Skyfire and KyaPay agreement

Skyfire launched the KyaPay Open Standard, the core innovation of which is a composite token:

• Kya token: contains identity information (such as “verified enterprise agent”).

• Pay token: includes the ability to pay (such as “pre-authorized 10 USDC”).

• kya+pay token: packages identity and payment, allowing agents to complete “guest checkout” without manually filling out forms.

2.3.2 Catena Labs and ACK (Agent Commerce Kit)

Catena Labs, founded by USDC architect Sean, has launched ACK to create “HTTP for smart commerce.” ACK emphasizes the use of W3C DID standards and account abstraction to allow agents to directly control on-chain smart contract wallets to achieve greater security than API keys.

2.3.3 Google AP2 and x402 extensions

Agent Payments Protocol (AP2), launched by Google, leverages “power of attorney” management rights and collaborated with Coinbase to develop an AP2 x402 extension to directly integrate cryptographic payment standards into the protocol.

2.4 Agent Credit Scoring and Risk Control

KYA is also the beginning of a reputation system.

• On-chain reputation (ERC-7007): Through ERC-7007 (a verifiable AI-generated content token standard), every successful agent interaction (such as paying on time and generating high-quality code) can be recorded on the chain to form a verifiable history.

• Real-time fusing: Financial institutions are deploying AI gateways. If a transaction agent's behavior deviates from the benchmark (such as high-frequency abnormal transactions), the system can immediately revoke their VC, triggering “digital suppression.”

3. Economic Restructuring: Addressing the “Invisible Tax” of Open Networks

A16z's Liz pointed out that AI agents are levying an “invisible tax” on open networks: agents extract large-scale data (context layer) from content websites in order to serve users, yet they are systematically bypassing the advertising and subscription models that support the production of such content. If this parasitic relationship is not resolved, it will cause the content ecosystem to dry up.

3.1 The “big decoupling”: the full arrival of the zero-click economy

In 2025, the digital publishing industry saw a “big decoupling”: searches increased, but clicks to websites fell precipitously.

3.1.1 Cruel data on traffic erosion

• Zero click rate spike: a16z predicts that traditional search engine traffic will drop 25% by 2026. According to Similarweb data, zero-click search rates rose to 65% in 2025.

• Click-through rate (CTR) collapse: DMG Media reports that when the AI Overview appeared above search results, the click-through rate for its content plummeted 89%. Even the top search results lost 34.5% of hits in front of AI abstracts.

3.2 Getting Rid of Static Licensing: A New Pay-As-You-Go Model

In response to this crisis, the industry is shifting from static annual data licensing (such as Reddit's deal with OpenAI) to usage-based compensation.

3.2.1 Perplexity's Comet Plus model

Perplexity AI's Comet Plus program is a typical early attempt:

• Mechanism: Establish an initial revenue pool of $42.5 million. When AI agents reference publisher content in their responses or visit pages on behalf of users, they trigger revenue distribution.

• Share: Publishers can receive up to 80% of the relevant revenue pool. This acknowledges the commercial value of “machine access.”

3.3 Technical standards: nanopayments and microattribution

In order to extend compensation to the entire network, a series of open technology standards are being implemented.

3.3.1 Nanopay and x402 Agreements

The HTTP 402 status code has finally been activated. The x402 protocol establishes a standard for “machine-native payments.”

• Workflow: Agent requests resources -> server returns 402 Payment Required and price (such as 0.001 USDC) -> Agent automatically signs payment via L2 blockchain (such as Base, Solana) or Lightning Network -> releases data after server verification.

• Economy: Traditional payment gateways can't handle transactions that cost a few pennies, while x402 combined with a low rate chain reduces costs to negligible amounts, making payment possible.

3.3.2 Machine Readable Rights: TdmRep and C2PA

• TDMRep (Text Data Mining Reservation Protocol): W3C community standard that allows websites to state in robots.txt or HTTP headers: “TDM rights reserved, pay/license required.” This provides a clear binary signal for the agent.

• C2PA (Content Source and Authenticity Alliance): Prove the original origin of content by embedding tamper-proof “content credentials”. Even when content is ingested by AI, the cryptographic signature provided by C2PA ensures that the attribution link continues to break, providing a basis for royalty distribution.

3.4 On-Chain IP Ownership: Story Protocol

A more radical change would be to tokenize the intellectual property itself. Story Protocol is dedicated to building a “programmable IP” layer.

• Mechanism: Creators register their work as an “IP asset” on Story Network.

• Automated licensing: The asset comes with a “Programmable IP License.” When the AI agent uses the data, the smart contract automatically enforces license terms (such as “5% royalty for commercial use”) and automatically distributes the proceeds. This has created a highly liquid IP market without the involvement of lawyers.

3.5 Looking Ahead: From SEO to AEO

By 2026, the marketing focus will shift from SEO to AEO or GEO.

• The goal: stop looking for “number one search rankings,” but rather pursue being “quoted” ** by AI or being the “preferred data source” for its reasoning process.

• Sponsorship context: The future advertising model will be “contextual injection.” Brand bidding enters the intelligence chain. For example, having a travel agent “remember” a hotel when planning an itinerary is the best option.

4. conclusions

The 2026 technology picture clearly shows that the friction between human-centered internet infrastructure and the demand for machine-centralization is forcing a complete reconstruction of the digital world.

1. Research paradigm: AI moves from assistance to autonomy. The AWA architecture allows AI to mass-produce scientific discoveries at low cost and transform “illusion” into creativity.

2. Identity systems: KYA has become the new frontier of financial compliance, empowering billions of AI agents with legal economic identities to safely navigate value networks.

3. Economic model: The network economy is shifting from an attention-based advertising model to a value-based payment and programmable IP model. x402, TDMRep, and Story Protocol form the railroad tracks of the new economy, solving the “hidden tax” problem and ensuring that data producers remain profitable in the zero-click era.

We're seeing the birth of an agent economy — an economy where software not only helps us work; they themselves are producers, consumers, and traders.