Sequoia Capital: 2026 will be the year of AGI, and programming agents have already fired the first shot!

General Artificial Intelligence (AGI) is no longer a distant future but has become a reality with the emergence of "Long-horizon agents." According to an article titled "2026: This is AGI" published on the 14th by Sequoia Capital partners Pat Grady and Sonya Huang, although there are still differences in the technical definition of AGI, from a functional perspective, artificial intelligence capable of autonomously solving problems has officially landed, and 2026 will be its year. Sequoia Capital states that coding agents are the first instance of AGI realization, and more types of agents are emerging. Unlike early conversational AI, the new generation of long-horizon agents can reason based on baseline knowledge like humans and achieve goals through continuous self-iteration. This leap in capability marks the transformation of artificial intelligence from a mere "conversationalist" to an "executor" that can actually deliver work. This shift will have profound implications for the business and investment sectors. Sequoia Capital analyzes that with the exponential growth of agent capabilities, the logic for founders to build products will fundamentally change—from selling software to directly "selling work outcomes." Future AI applications will no longer be just auxiliary tools but entities that can work alongside as "colleagues" around the clock, with users transitioning from independent contributors to managers of agent teams. As Claude Code and other coding agents recently crossed critical capability thresholds, market perceptions of AGI have been reshaped. The article emphasizes that through reinforcement learning and optimization of agent architectures, the ability of agents to handle complex tasks is growing at a rate that doubles every seven months, which will fundamentally change the talent structure and productivity boundaries of enterprises. Functional Definition: AGI is the Ability to "Solve Problems Autonomously" Sequoia Capital states that as investors, they do not intend to engage in the technical definition debate of AGI but propose a pragmatic functional definition: AGI is the "ability to solve problems autonomously." For businesses that want to succeed, how AI achieves its goals is not important; what matters is whether it can truly complete tasks. The article breaks down AI with this capability into three core elements: Baseline Knowledge (Pre-training): This is the core driving force of the 2022 ChatGPT moment. Reasoning Ability (Inference Calculation): Achieved with the release of the o1 model by the end of 2024. Iterative Ability (Long-horizon Agents): This is the latest breakthrough, where AI can autonomously work like a general intelligent human, correct errors within hours, and decide on the next steps without specific instructions. From Instructions to Autonomy: The Work Loop of Agents To illustrate what "solving problems autonomously" means, the article uses a recruitment scenario as an example: when a founder needs to find a developer relations head who understands technology and is active on social media, the traditional approach is to post a job description. In contrast, agents can autonomously execute complex search loops According to the article, intelligent agents can complete the psychological cycle of human recruitment experts in 31 minutes: They not only search for relevant positions at competing companies like Datadog and Temporal on LinkedIn but also turn to YouTube to filter high-engagement speakers and further cross-reference activity levels and content quality on Twitter. The agents can even keenly capture potential resignation signals by analyzing the decline in posting frequency, ultimately selecting the best candidates and drafting personalized outreach emails. This ability to establish hypotheses, test, iterate, and adjust direction until goals are achieved in ambiguous environments is the core characteristic of long-term intelligent agents. Although they still produce hallucinations or lose direction, their developmental trajectory is irreversible, and errors are becoming increasingly correctable. Technical Path: Dual Drive of Reinforcement Learning and Agent Architecture Regarding how to achieve this leap, Sequoia Capital points out that getting models to think for extended periods is no easy task. Currently, two technical paths have been proven effective and scalable: First is Reinforcement Learning, primarily led by research laboratories. Through continuous "nudging" and guidance during training, models are taught to maintain focus over long periods. Significant progress has been made in multi-agent systems and tool reliability. Second is Agent Harnesses, which falls under the application layer. Developers design specific scaffolding (such as memory handoff, compression, etc.) to circumvent known limitations of the models. Currently well-received products in the market, such as Manus, Claude Code, and Factory’s Droids, benefit from their excellent architectural design. According to METR's tracking of AI's ability to complete long-term tasks, progress in this field is growing exponentially. Based on current trends, intelligent agents will reliably complete tasks that human experts currently take a whole day to accomplish by 2028, and by 2034, they will be able to handle a year's worth of work. Business Transformation: From Software to "Digital Employees" "Can you hire an intelligent agent?" Sequoia Capital believes this is the litmus test for AGI. The current market landscape indicates that specialized intelligent agents are rapidly emerging across various industries, from OpenEvidence in pharmaceuticals, Harvey in law, to XBOW in cybersecurity. This signifies a massive paradigm shift for entrepreneurs. AI applications in 2023 and 2024 are mostly "conversational agents" with limited impact; whereas applications from 2026 onwards will be "executors." This shift makes "sales work" possible. Founders need to rethink: which ongoing tasks that require continuous attention can be taken over by intelligent agents? How can pricing and packaging be oriented towards "outcomes" rather than "tools"? The article concludes by urging the market to "saddle up" for the exponential growth of long-term intelligent agents. While today's intelligent agents may only reliably work for about 30 minutes, they will soon be able to handle a full day's workload, and ultimately even manage tasks equivalent to a century's worth of human work This means that what was once considered an overly ambitious roadmap—such as cross-referencing 200,000 clinical trial data or reconstructing the entire U.S. tax code—has now become feasible. In the year of AGI, ambitious plans are gradually transforming into realistic business strategies