Historical Behavior

Andrew Chen's historical behavior in AI investment shows a "paradigm shift" characteristic. In 2023, he invested in the "Copilot" model (Jasper, Runway), in 2024 he shifted to the "Agent" model (Devin, Cognition), and in 2026 he explicitly proposed the view of "action > assistance." Compared to Sam Altman (OpenAI CEO, who proposed the "Copilot era" in 2023 and launched the "Operator agent" in 2026) and Marc Andreessen (a16z partner, who proposed "AI devours software" in 2024 and "agents devour workflows" in 2026), Chen's perspective is more "user-centered"—shifting from "AI assisting humans" to "AI replacing humans."

Capital Path

The capital impact of agents is clear: in 2026, corporate AI spending will reach $200 billion, with the "Copilot" model accounting for 60% ($120 billion, $50 per month subscription per user) and the "Agent" model accounting for 40% ($80 billion, pay-per-result, $10-100 per task). If in 2027 the Agent model accounts for 70% and the Copilot model for 30%, the structure of corporate AI spending will be restructured: from "subscription fees" (fixed costs) to "pay-per-result" (variable costs). Funding is shifting from "tool procurement" (one-time purchase of Copilot) to "outsourcing services" (pay-per-task Agent), reflecting the paradigm of "AI as employees." Devin will charge $500 per month in 2026 (replacing one junior engineer with an annual salary of $100,000), with an ROI of about 20:1.

Analogy + Industry Position

Similar cases: In the 2010s, "SaaS vs On-Premises"—SaaS accounted for 20% of enterprise software in 2010, and is expected to account for 80% in 2026, replacing one-time purchases with subscription payments. In the 2020s, "Gig Economy vs Full-Time Employment"—Upwork accounted for 10% of the freelance market in 2020, and is expected to account for 30% in 2026, replacing fixed salaries with pay-per-task. Currently, AI agents are in the "Agent explosion phase"—compared to 2023 (Copilot dominated, 90% market), and 2026 (Agent rising, 40% market), it is expected that by 2030, Agents will account for 80% and Copilots 20%. The industry position is at the "AI autonomy inflection point"—users are shifting from "AI help me write code" (Copilot) to "AI help me write complete code and deploy" (Agent), and from "AI help me analyze data" (Copilot) to "AI help me analyze and generate reports and send emails" (Agent).

Structural Judgment

Essentially, this is a technological substitution leading to workflow reconstruction. Traditional Copilot involves "humans in the loop"—AI generates content, human reviews, modifies, and publishes, improving efficiency by 2-5 times. Agents involve "humans on the loop"—AI generates content, automatically reviews, modifies, and publishes, with human intervention only in exceptional cases, improving efficiency by 10-50 times. This explains why Chen proposed "more action, less co-pilot"—it is not a "function upgrade," but a "paradigm shift." When Agents can autonomously complete 90% of tasks (writing code, analyzing data, sending emails), humans only need to handle 10% of exceptions, making the "review bottleneck" the new constraint. This drives the "polarization of workflows"—high-end workflows (complex decisions, high risk) retain human review, low-end workflows (repetitive, low risk) become fully automated, and the middle layer (semi-automated) disappears. As Agents become widespread, "humans in the loop" will shift from "default mode" to "exception mode," transitioning workflows from "AI assisting humans" to "humans supervising AI."

ABAB News · Law of Cognition

Co-pilot is assistance, agent is replacement
Humans in the loop are a bottleneck, humans on the loop are the future
Pay-per-result is the ultimate goal of AI commercialization