OpenAI's Case for Agents: Work Is Shifting from Prompts to Delegation

What it is / 这是什么

OpenAI has published How agents are transforming work, a report making a blunt argument: the interesting shift in AI is no longer better answers, it's delegation. An agent, in OpenAI's framing, is a system that takes a goal rather than a prompt — it plans the steps, calls tools (a browser, a code runtime, internal APIs), executes them, and comes back with a finished result instead of a suggestion.

That's a different product category from the chatbot era. ChatGPT answering questions is assistance; an agent opening your CRM, reconciling records, and drafting the follow-up emails is execution.

Why it matters / 为什么重要

The unit of work changes. With chat, a human decomposes the task, prompts step by step, and stitches results together — the human is still the workflow engine. With agents, you hand over the whole task and review the output. That collapses the cost of the boring middle: research compilation, data cleanup, first-draft code, QA passes.

The HN-flavored take: this is the RPA promise, except it might actually work this time, because agents handle ambiguity that brittle click-recording bots never could. The failure mode also changes — instead of a wrong answer you can ignore, you get wrong actions, which is why every serious agent deployment described in OpenAI's report keeps humans at approval checkpoints.

Key features / 核心特性

What agentic work looks like / 智能体工作的形态

• Goal-level delegation: you specify the outcome and constraints, not the steps.
• Tool use: browsing, code execution, file manipulation, and API calls — the same surface a human employee touches.
• Long-horizon tasks: minutes to hours of autonomous work, not single-turn replies.
• Human checkpoints: review gates before irreversible actions (sending, deploying, paying).

OpenAI's own product line maps to this: deep research for multi-source analysis, Codex for delegated software tasks, and AgentKit for teams building custom agents on their internal tools.

How it compares / 横向对比

Against Anthropic's Claude Code and Agent SDK, the framing is nearly identical — both bet that agentic coding is the beachhead because code is verifiable (tests either pass or they don't). Against classic RPA (UiPath-style automation), agents trade deterministic replay for adaptability. Against plain chatbots, agents cost more per task but replace the whole loop, not one turn of it. The honest current state: agents are strongest where output is checkable and weakest where judgment is the product.

Who should use it / 适合谁

Developers automating their own toil (code review prep, dependency upgrades, test triage) get value today. Ops, support, and research teams with repetitive multi-step workflows are next. If your work is mostly novel judgment calls with no verification loop, agents assist but don't yet replace — read the report and start with one bounded workflow rather than a moonshot.

FAQ / 常见问题

Q: What separates an agent from a chatbot? A chatbot answers; an agent plans, acts through tools, and returns finished work.

Q: Is it safe to let agents act autonomously? Only with checkpoints. Gate anything irreversible behind human approval.

Q: Where's the fastest ROI? Verifiable, repetitive multi-step tasks — code with tests, research with citations, structured data work.