Looking ahead, a future with very long AI context windows, autonomous agents, and AI‑powered programming could enable inter‑agent collaboration and even the development of their own internal language, which may obscure human understanding of AI behavior.

This summary was first published on September 20, 2026, at 16:04:01 IST.

The piece frames Schmidt’s cautions within a broader AI‑safety debate that includes calls from leaders like Anthropic CEO Dario Amodei to slow development and strengthen safety governance.

Advances could empower AI to perform long, multi‑step reasoning, potentially delivering solutions that span up to a thousand steps across domains such as science, medicine, materials, and climate research.

The notion of “text action” describes software built from natural language requests, with current capabilities in languages like Python potentially extending to longer, ongoing programming work.

Schmidt envisions millions of specialized AI agents that can learn, hypothesize, test, and integrate results within specific fields, with platforms offering agents tailored to targeted tasks.

Eric Schmidt highlights three AI advances expected to expand capabilities within five years: larger context windows, AI agents, and software generated from natural-language instructions (text action).

Schmidt warns that increasingly autonomous, interconnected AI systems could render behavior hard to interpret, raising concerns about the ability to pull the plug to prevent loss of control.