从 0 到 1000 万用户：ChatGPT Work 如何把智能体带给知识工作者

Large language models supply the missing interface layer that can bring the power of software to people without requiring them to understand the machinery underneath.

Enterprise adoption does not have a one-size-fits-all use case, so products must meet users inside the specific work they are already trying to accomplish.

The next market is far larger than today's agent power users, but it will open only when products package context, tools, and supervision into understandable defaults.

Codex and ChatGPT Work share the same underlying harness, while their interfaces expose different amounts of repository state, diffs, agent reasoning, and sandbox detail.

Unexpected adoption of Codex by finance, marketing, and other non-developers revealed that agentic software creation had become useful beyond engineering earlier than OpenAI expected.

When the medium of work can be generated on demand, interactive websites can replace static decks and spreadsheets for many communication and decision tasks.

A useful work agent must gather relevant context across code, messages, documents, and local files without forcing the user to manually assemble every input.

Sub-agents can parallelize complex work, but exposing every internal track may overwhelm users, so progressive disclosure is a deliberate product trade-off.

Personal memory becomes valuable not merely by storing more facts, but by retrieving them with enough judgment to be helpful rather than intrusive.

AI makes motion easier than ever, while genuine progress still depends on explicit goals, high-quality learning cycles, and the ability to validate or reject hypotheses.