With the rapid evolution of inference model capabilities in the second half of 2025, the competitive focus of AI in 2026 is shifting from being smarter to being more capable, trustworthy, and accurately valued. Three members of the a16z crypto research and investment team recently presented their predictions for three major trends in AI development in 2026, focusing on research workflows, agent identity infrastructure, and open network economic models.
Harvard professor Scott Kominers predicts that by 2026, AI will have moved from being an assistant to a research partner, offering creative insights. He stated that by November 2025, he was already able to interact with models using abstract commands similar to those given to a PhD student, and obtain novel answers.
Circle co-founder Sean Neville believes that enabling AI agents to transact as entities, with traceability, authorization, and verifiability through KYA (Know Your Agent) will be a key trend.
Liz Harkavy of the a16z crypto investment team believes that AI extracts content from the internet and grows itself without contributing advertising traffic, leading to a significant mismatch between the contextual and execution layers of the internet. She advocates that rewards should be distributed to each entity that contributed information, data, or content when an agent's task is successful. She also mentioned that blockchain-supported nanopayments and more mature attribution standards may be a viable technological path.
Trend 1: AI is evolving from an assistant to a research partner, becoming more capable of undertaking substantive research tasks.
Scott Kominers, a member of the a16z crypto research team and a professor at Harvard Business School, said that it would still be quite difficult for consumer-grade AI models to understand his research workflow in early 2025, but by November 2025, he was able to interact with the models using abstract instructions that were similar to those used to guide doctoral students, and the models would sometimes return novel and correctly executed answers.
A new, erudite research style will emerge in AI in 2026.
Kominers points out that the use of AI in research is becoming more widespread, especially in disciplines that require reasoning. Models are beginning to directly assist in exploration and can even automatically solve extremely difficult math competition problems like Putnam. However, which disciplines will benefit the most and in what ways remain unanswered questions.
However, he anticipates a new polymath research style emerging in 2026: researchers will place greater emphasis on proposing conjectures that link concepts and quickly extrapolating verifiable directions from still speculative answers.
AI research and evolution still carries the risk of illusion; encryption technology can offer assistance.
He also admitted that this research method inevitably carries the risk of inaccuracy and illusion, but when the model is smart enough, giving it room for abstraction and allowing it to diverge may, like human creativity, occasionally lead to breakthroughs. He proposed that by 2026, the research AI workflow will be more like agent-wrapping-agent: using multiple layers of models to evaluate, verify, and then synthesize the conclusions.
However, Kominers also cautioned that running such inference agent clusters on a large scale requires better model interoperability and methods to identify and appropriately compensate for the contributions of each model. He believes that encryption technology could help address these two issues.
Trend Two: From KYC to KYA – Knowing Your Agent Becomes a Bottleneck in the Agent Economy
Sean Neville, co-founder of Circle, architect of USDC, and current CEO of Catena Labs, focuses on a key bottleneck in the agent economy: the shift from intelligence to identity.
Neville points out that in sectors like financial services, the number of non-human identities far exceeds that of human employees, reaching as high as 96 to 1. However, most of these identities remain ghosts, unable to open accounts or assume responsibility. Therefore, he argues that the next key primitive is KYA (Know Your Agent).
By definition, KYA addresses the requirement that an agent must possess verifiable, traceable, and accountable credentials to conduct transactions on behalf of an entity. Just as humans need credit scores for lending and borrowing, agents also need encrypted, signed credentials to link their principal, constraints, and liability. Before KYA was implemented, merchants and service providers would typically block agents' access at the firewall level to prevent fraud, abuse, and unclear liability.
He also stated frankly that the KYC industry and regulatory framework, which took decades to build, may now only have a few months to explore and implement KYA.
Trend 3: AI proxies impose hidden taxes on the internet, extracting content value and circumventing revenue generation.
Liz Harkavy of the a16z crypto investment team focuses on how agents are reshaping the economic foundation of the open web. She describes the rise of AI agents as imposing an invisible tax on the open web: agents extract content from ad-supported websites (which she calls the context layer) to provide users with more convenient answers and actions (the execution layer), but systematically bypass the revenue streams that support content production, such as ad impressions, subscription conversions, and referrals.
Harkavy argues that this leads to a significant mismatch between the contextual and implementation layers of the internet: content providers bear the costs, agents and platforms absorb the value, and the original monetization path is cut off. She points out that current AI licensing deals are mostly just a smokescreen, often compensating content providers with a small portion of the lost traffic revenue, which may still be financially unsustainable in the long run.
She argues that to prevent the open internet from being hollowed out (while also protecting the diverse content sources upon which AI relies), a large-scale deployment of technological and economic solutions is needed by 2026: such as next-generation sponsored content models, micro-attribution systems, or other new funding models. The key shift lies in moving from static licensing to real-time, usage-based compensation, allowing value to flow automatically.
Upon successful completion of an agent's task, the reward will be distributed to each entity that contributed information, data, or content. She also mentioned that blockchain-supported nanopayments, along with more mature attribution standards, may be one viable technological path.
This article, a16z's "Three Predictions for AI in 2026: The Rise of Research-Oriented AI, KYA Taking Over from KYC, and the Crisis of Hidden Taxes on the Internet," first appeared on ABMedia, a ABMedia .
