Pat Grady is a long-tenured Sequoia partner whose public record spans growth-stage enterprise software and newer AI applications. Since November 2025, he has shared stewardship of Sequoia with Alfred Lin. His current thesis is that long-horizon agents can become a new unit of software value. That is an investor’s forecast, not a settled definition of AGI or proof that agent businesses already have durable economics.

As of September 13, 2026, Sequoia’s profile lists Grady’s current and historical company relationships, including ServiceNow, Zoom, Snowflake, Harvey, OpenAI, and several agent startups. The page is the firm’s own portfolio record. It should not be converted into a personal “$250 billion portfolio” because market values, ownership stakes, realized proceeds, and individual attribution are different quantities.

The short answer: Grady is betting that software will sell completed work

Traditional enterprise software sells tools and seats. The agent thesis says some products can instead accept a goal, use tools, iterate, and deliver a result. If that transition holds, product design, pricing, and procurement all change:

  • Interfaces move from navigation toward delegation and review.
  • Reliability is measured across a task sequence, not one model response.
  • Pricing can move from seats toward usage or accepted outcomes.
  • Security expands from data access to permissioned action.
  • Gross margin includes inference, retries, monitoring, and human repair.

Grady and Sonya Huang made the strongest version of this claim in “2026: This is AGI”. They use a functional definition—an agent that can “figure things out” over longer horizons—and argue that coding agents show the direction of travel. The essay is clearly labeled investor analysis. Its examples and extrapolations are not independent measurements of every cited product.

Enterprise software provides the relevant discipline

Grady’s earlier portfolio history matters less as a list of famous logos than as exposure to recurring enterprise problems: long sales cycles, integration, security reviews, customer concentration, implementation work, and expansion within accounts. AI applications do not escape those constraints.

Sequoia’s page on ServiceNow’s history includes named participants discussing product scope, leadership transitions, culture, and scaling. It is a firm-produced oral history, so it reflects Sequoia’s framing. It nonetheless illustrates the difference between a durable workflow platform and a feature that is easy for another vendor to reproduce.

That distinction is central to AI. A model wrapper can demonstrate value quickly, yet lose differentiation when model providers add the same feature. A more defensible application owns workflow context, integrations, permissions, evaluation data, and the feedback loop that improves accepted outcomes.

The evidence behind longer-horizon agents is narrower than the slogan

Grady and Huang cite research from METR on the length of software tasks agents can complete. METR’s methodology and results define a 50% task-completion time horizon using human completion time for a suite of software and research tasks. METR found a strong historical improvement trend, while warning that extrapolation assumes the trend continues.

The boundary matters. A longer horizon on scorable software tasks does not automatically mean reliable operation in healthcare, law, recruiting, finance, or physical systems. Real work often has incomplete specifications, changing goals, social coordination, irreversible effects, and weak feedback. A model can also complete a long task occasionally while remaining unsuitable for unattended production.

The responsible interpretation is that agent persistence is improving on tested tasks. The commercial hypothesis—that this improvement creates dependable, high-margin businesses across domains—still requires product-level evidence.

Harvey shows the application-layer bet

Sequoia says it partnered with legal AI company Harvey in 2023. Its Harvey company page names Grady among the relevant partners and describes the company as serving professional-services work. This establishes Sequoia’s relationship and positioning; it does not independently validate accuracy, customer outcomes, valuation, or revenue.

Legal work illustrates why application context matters. A useful system must retrieve the right matter documents, respect ethical walls, preserve citations, expose uncertainty, and fit review workflows. Model quality is necessary but not sufficient. The product’s durable value lies in reducing accepted-work time without weakening confidentiality or professional accountability.

The same logic applies to other vertical agents. A buyer should evaluate the full path from request to approved outcome, not a selected transcript.

A practical agent-company scorecard

Founders and buyers can test the Grady thesis with a shared scorecard:

  1. Accepted completion rate: what fraction of representative tasks pass expert review without material repair?
  2. Review burden: how much expert time is required per accepted result?
  3. Long-sequence reliability: how quickly does success fall as tools, steps, and elapsed time increase?
  4. Action safety: are permissions, approvals, spend limits, and reversibility enforced outside the model?
  5. Unit economics: what is gross margin after inference, retries, observability, and human operations?
  6. Retention: do customers expand after the pilot, and are they using the product for core work?
  7. Supplier resilience: what breaks when the underlying model, price, or policy changes?

These measures protect against two errors: dismissing a real capability shift because early agents fail, and declaring a new software era from a curated demo.

Stewardship adds a second accountability layer

The co-steward role concerns the institution as well as individual deals. Grady and Lin must allocate attention across stages and sectors, develop partners, manage portfolio conflicts, and communicate risk to limited partners. Public material does not provide a full map of their decision rights, so claims that either person alone controls Sequoia’s AI strategy go beyond the evidence.

The firm’s AI essays also create an important disclosure issue. When Sequoia discusses a market that includes its portfolio companies, readers should separate market evidence from portfolio advocacy. The firm commonly discloses its relationships, but disclosure does not make projections independent.

What remains unknown

Public sources do not reveal Grady’s personal wealth, carried interest, deal-level returns, confidential board advice, or the internal attribution for Sequoia investments. They also do not establish that long-horizon agents are equivalent to AGI under scientific or policy definitions. Those are not presented as facts here.

The evidence supports a narrower conclusion: Grady has applied enterprise-software investing experience to a thesis that AI products will increasingly deliver work rather than merely expose tools. The thesis is plausible and testable. Its success depends on reliability, control, retention, and economics—not on a portfolio’s aggregate market capitalization or an expansive label.

Source and correction note

This revision removes an unsupported “$250 billion-plus portfolio” figure, personal return estimates, and invented accounts of investment discussions. Sequoia claims and forecasts are labeled as such; METR’s underlying evidence and limitations are linked separately. Roles and public materials were checked through September 13, 2026.