Long-horizon AI agents and Sequoia's case that AGI is here

On January 14, 2026, Sequoia Capital partners Pat Grady and Sonya Huang published "2026, This is AGI." The essay makes one central claim. Long-horizon agents are functionally AGI, and 2026 will be their year. Sequoia is explicit that this is an investor's definition, built on real-world impact rather than a technical benchmark.

Sequoia's functional definition of AGI

Sequoia defines AGI as "the ability to figure things out." The authors break that ability into three parts and map each to a technical shift:

• Baseline knowledge, from pre-training. This drove the original ChatGPT moment in 2022.
• Reasoning over that knowledge, from inference-time compute. This arrived with OpenAI's o1 in late 2024.
• Iteration toward an answer, from long-horizon agents. Sequoia says coding agents such as Claude Code crossed this threshold in the weeks before publication.

The point is that agents can now work for extended periods, make and fix their own mistakes, and decide what to do next without step-by-step instructions.

The 31-minute recruiting example

The essay illustrates the idea with a hiring task. A founder asks an agent to find a developer relations lead. The agent searches job titles, pivots to conference talks, cross-checks social activity, filters for signs that a candidate may be open to a move and drafts a personal outreach email. Sequoia reports the whole loop took 31 minutes and produced a shortlist of one. The authors also state plainly that agents still hallucinate, lose context and sometimes pursue the wrong path.

The exponential Sequoia is betting on

Sequoia points to research group METR, which tracks how long a task AI can complete. The essay says that length has been doubling roughly every seven months. Extending that trend, Sequoia projects agents could reliably complete tasks that take a human expert a full day by 2028 and a full year by 2034. The authors estimate that agents today work reliably for about 30 minutes.

Two technical approaches are driving progress, according to the essay. Research labs use reinforcement learning to keep models on task for longer. Application companies build agent harnesses, the scaffolding that manages memory hand-offs and context. Sequoia names Manus, Claude Code and Factory's Droids as products known for strong harnesses.

From talkers to doers

Sequoia draws a line between AI applications of 2023 and 2024, which it calls talkers, and those of 2026 and 2027, which it calls doers. Doers will be used all day, with several running in parallel. The essay says users will move from working as individual contributors to managing teams of agents.

The authors list companies they see functioning as hired roles. Harvey's agents act as a law associate. XBOW acts as a penetration tester. Juicebox acts as a recruiter. Day AI covers several go-to-market roles.

Four questions Sequoia puts to founders

The essay closes with a checklist, which PacTech restates here as instructions:

• Define the work. Identify tasks in your domain where sustained attention is the bottleneck.
• Productize the work. Plan how your interface moves from a chat box to delegation.
• Make the work reliable. Invest in your agent harness and build a tight feedback loop.
• Sell the work. Price and package against value and outcomes, not seats.

What this means for early-stage teams

The essay's most useful line for a seed-stage founder is practical. If agents can complete half-hour tasks reliably today and the horizon keeps doubling, a roadmap built for today's limits will be out of date within a year. Build the product for the horizon you expect at launch, and test it against today's failure modes. Sequoia's closing claim is that the ambitious version of a roadmap is now the realistic one.

Sources

1. 2026, This is AGI
   Sequoia Capital (Pat Grady, Sonya Huang) · Jan 14, 2026