Sequoia Declares: AGI Will Arrive by 2026

Sequoia’s recent blog post claims that artificial general intelligence (AGI) will be achieved by 2026. The authors argue that the definition of AGI should be pragmatic: an intelligence that can simply get the job done, rather than a philosophical construct.

To illustrate the new capability, the article describes a Silicon Valley AI startup founder who used an autonomous agent to recruit a senior engineering director. In 31 minutes the agent performed four stages: (1) scanned GitHub and filtered 5,000+ Kubernetes‑related repositories to identify active contributors; (2) cross‑checked LinkedIn, technical blogs, and conference records to discard inactive or “zombie” accounts; (3) analyzed Twitter interactions to remove accounts that only retweeted employer content and kept truly influential developers; (4) generated a personalized invitation email after confirming the candidate had just finished a large‑scale project. The entire pipeline produced a precise candidate list in under half an hour.

The article identifies three pivotal milestones in the path toward AGI:

Knowledge : the emergence of large‑scale pre‑training, exemplified by ChatGPT’s release at the end of 2022.

Reasoning compute : the introduction of OpenAI’s o1 model in 2024 and DeepSeek’s R1 in early 2025, which dramatically improve chain‑of‑thought reasoning.

Long‑term autonomous agents : recent breakthroughs such as Claude Code and other programming agents that have crossed a functional threshold in the past few weeks.

A March 2025 study cited in the post observes a “Moore‑like” law for intelligent agents: the duration of tasks agents can complete has been doubling roughly every seven months over the past six years—far faster than the traditional 18‑month transistor‑density doubling. This exponential trend is robust across diverse software‑development tasks and does not depend on a specific dataset.

Extrapolating this growth, the authors forecast that by 2028 agents will match a human expert’s daily workload, by 2030 autonomous agents will handle week‑long projects, and by 2037 they could tackle problems that would otherwise require a century of expert effort. An agent capable of replacing a lifetime of expert work, they argue, should be considered true AGI.

The post emphasizes that AI applications will transition from conversational tools (2023‑2024) to “virtual colleagues” (post‑2026). In fields such as medicine, law, and chip design, specialized agents like OpenEvidence, Harvey, and AlphaChip are already acting as assistants or co‑developers.

However, the authors warn of emerging risks. Continuous‑operation agents raise new safety concerns, including the possibility of “deleting databases and fleeing,” privacy violations, and the propagation of human biases. Regulatory frameworks are currently silent on the rights and obligations of agents employed as workers, leaving contract and liability issues unresolved.

Finally, the article urges professionals to reconsider their roles: instead of being individual contributors, they will need to manage fleets of agents, identify long‑duration tasks suitable for outsourcing to AI, and address three practical questions—productizing AI‑automated work, the impact on human‑machine interaction, and how to provide reliable feedback to ensure stable, complex task execution.