Key Takeaways
- • Sequoia Capital argues in its 2026 essay that AGI is already here, defining it functionally as "the ability to figure things out" rather than through a technical benchmark.
- • The firm identifies three ingredients: baseline knowledge from pre-training (the ChatGPT moment in 2022), reasoning via inference-time compute (o1 in late 2024), and iteration through long-horizon agents (Claude Code and other coding agents crossing a capability threshold in the weeks before publication).
- • Citing METR tracking, Sequoia says the length of tasks AI agents can complete reliably doubles roughly every seven months, implying day-long expert tasks by 2028, year-long by 2034, and century-long by 2037.
- • Sequoia says 2026 will be the year of long-horizon agents and that AI applications will shift from "talkers" (2023-2024) to "doers" that operate like colleagues and can be sold as completed work.
- • The essay concedes agents still hallucinate, lose context, and sometimes pursue wrong paths, but argues these failures are increasingly fixable.
Summary
Sequoia Capital has published an essay titled "2026: This is AGI," arguing that artificial general intelligence is present now rather than a distant goal. The firm acknowledges that AI researchers have never settled on a concrete definition of AGI, and states that it lacks the authority to propose a technical one. Instead, Sequoia offers a functional definition: AGI is "the ability to figure things out," judged by real-world impact rather than technical capability.
The essay frames this ability as the combination of three ingredients: baseline knowledge gained through pre-training, which fueled the original ChatGPT release in 2022; reasoning over that knowledge through inference-time compute, which arrived with OpenAI's o1 in late 2024; and iteration toward answers through long-horizon agents, which Sequoia says crossed a capability threshold in the weeks before publication with Claude Code and other coding agents. Sequoia illustrates this with an example of an agent that identifies a developer relations candidate in 31 minutes through successive searches, filtering and hypothesis testing.
Sequoia cites METR's tracking of AI performance on long-horizon tasks, which it says doubles roughly every seven months. On that trajectory, the firm projects that agents will reliably complete tasks requiring human experts a full day by 2028, a full year by 2034, and a full century by 2037. The essay credits two technical approaches for this progress: reinforcement learning, pursued by research labs, and agent harnesses, the scaffolding built at the application layer, with products such as Manus, Claude Code and Factory's Droids cited as examples. Sequoia currently puts reliable agent operation at roughly 30 minutes.
The firm lists a range of emerging agent-based offerings, including OpenEvidence's Deep Consult in medicine, Harvey in law, XBOW in cybersecurity, Traversal in DevOps, Day AI in go-to-market functions, Juicebox in recruiting, and Harmonic's Aristotle in mathematics. Sequoia argues that applications from 2026 and 2027 will be "doers" rather than talkers, used all day in parallel and enabling workers to manage teams of agents, and it urges founders to consider how to productize and price work by outcome. The essay also concedes that agents still fail in ways that remain a risk for deployment.
The essay frames this ability as the combination of three ingredients: baseline knowledge gained through pre-training, which fueled the original ChatGPT release in 2022; reasoning over that knowledge through inference-time compute, which arrived with OpenAI's o1 in late 2024; and iteration toward answers through long-horizon agents, which Sequoia says crossed a capability threshold in the weeks before publication with Claude Code and other coding agents. Sequoia illustrates this with an example of an agent that identifies a developer relations candidate in 31 minutes through successive searches, filtering and hypothesis testing.
Sequoia cites METR's tracking of AI performance on long-horizon tasks, which it says doubles roughly every seven months. On that trajectory, the firm projects that agents will reliably complete tasks requiring human experts a full day by 2028, a full year by 2034, and a full century by 2037. The essay credits two technical approaches for this progress: reinforcement learning, pursued by research labs, and agent harnesses, the scaffolding built at the application layer, with products such as Manus, Claude Code and Factory's Droids cited as examples. Sequoia currently puts reliable agent operation at roughly 30 minutes.
The firm lists a range of emerging agent-based offerings, including OpenEvidence's Deep Consult in medicine, Harvey in law, XBOW in cybersecurity, Traversal in DevOps, Day AI in go-to-market functions, Juicebox in recruiting, and Harmonic's Aristotle in mathematics. Sequoia argues that applications from 2026 and 2027 will be "doers" rather than talkers, used all day in parallel and enabling workers to manage teams of agents, and it urges founders to consider how to productize and price work by outcome. The essay also concedes that agents still fail in ways that remain a risk for deployment.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.