Author:a16z crypto

Compiled by: Deep Tide TechFlow

This Year, AI Will Take on More Substantial Research Tasks

As a mathematical economist, back in January 2025, I struggled to get consumer-grade AI models to understand my workflow; however, by November 2025, I could give abstract instructions to AI models as I would to a PhD student... and sometimes they even returned novel and correct answers. Beyond my personal experience, AI is being more widely applied in research, especially in the field of reasoning. These models not only directly assist the discovery process but can also autonomously solve difficult problems like the Putnam exam (perhaps the world's most challenging university math test).

It's still uncertain in which areas this research assistance will be most helpful and how exactly it will be implemented. But I predict that this year, AI research will promote and reward a new style of "generalist" research: one that focuses more on conceptualizing the relationships between various ideas and can quickly infer from more hypothetical answers.

These answers may not be entirely accurate, but they can still guide research in the right direction (at least within a certain topological structure). Ironically, this is somewhat like harnessing the power of model "hallucination": when models are "smart enough," giving them abstract space to brainstorm may still produce some nonsensical results—but sometimes it leads to breakthrough discoveries, much like how humans can be most creative when not thinking linearly or following explicit directions.

Reasoning in this way requires a new style of AI workflow—not just simple "agent-to-agent" interactions, but a complex collaborative model of "nested agents." In this model, different layers of models assist researchers in evaluating the proposals of earlier models and gradually refining the essence. I already use this method to write papers, while others are using it for patent searches, inventing new forms of artistic works, and even (regrettably) discovering new smart contract attack methods.

However, operating these nested reasoning agent combinations for research still requires better interoperability between models and a way to identify and appropriately compensate each model's contributions—issues that blockchain technology may help solve.

—Scott Kominers(@skominers), a16z crypto research team member, Harvard Business School professor

From "Know Your Customer" (KYC) to "Know Your Agent" (KYA): The Shift in Identity Verification

The bottleneck in the agent economy is shifting from intelligence to identity verification. In financial services, the number of "non-human identities" now exceeds human employees by 96 times—yet these "identities" remain "ghosts" unable to access banking services.

The key missing infrastructure here is "Know Your Agent" (KYA). Just as humans need credit scores to obtain loans, agents need cryptographically signed credentials to conduct transactions—credentials link the agent to its principal, constraints, and liabilities. Until this infrastructure is established, merchants will continue to block these agents at the firewall.

The industry that built KYC infrastructure over the past decades now has only a few months to figure out how to implement KYA.

—Sean Neville(@psneville), Co-founder of Circle, Architect of USDC; CEO of Catena Labs

Solving the "Invisible Tax" on Open Networks: Economic Challenges in the AI Era

The rise of AI agents is imposing an "invisible tax" on open networks, fundamentally disrupting their economic foundation. This disruption stems from the growing mismatch between the internet's "context layer" and "execution layer": currently, AI agents extract data from ad-supported websites (the context layer) to provide convenience to users while systematically bypassing the revenue sources that support the content (such as ads and subscriptions).

To prevent the gradual decline of open networks (and protect the diverse content that fuels AI), we need to deploy technical and economic solutions on a large scale. These solutions may include next-generation sponsored content models, micro-attribution systems, or other new funding models. However, existing AI licensing agreements have proven financially unsustainable, often compensating content providers for only a fraction of the revenue lost due to AI traffic diversion.

The web urgently needs a new techno-economic model that allows value to flow automatically. The key shift in the coming year will be from static licensing models to compensation mechanisms based on real-time usage. This means testing and scaling systems—possibly leveraging blockchain-enabled nanopayments and complex attribution standards—to automatically reward every entity that contributes information to the successful completion of an AI agent's task.

—Liz Harkavy(@liz_harkavy), a16z crypto investment team