Over the past few years, there has been a very interesting automation tool called Zapier: it connects various services and automates tasks. The industry has also given it a cool name: RPA (Robotic Process Automation).
Looking at it today, isn't this exactly the "Agent" that everyone is hyping up?
Yes. For example, there's a daily task: summarize the daily performance of the US market and output it as a table.
OpenAI has turned this task into a simple instruction:
Very simple.
This task has been running for two days now (as the feature rollout is less than three days old).
Here are the results from the first day; the content is there, but it's not very user-friendly.
So, let's modify the prompt.
Today's result has been optimized.
OpenAI is increasingly making no secret of its ambition to unify AI applications, and RPA is becoming the first casualty on this path. However, extending more functionality must also break through the "cages" set by operating systems for web applications or apps:
OpenAI's Agent
Daoming, WeChat Official Account: Daoming Digital Nomad Life
After the release of ChatGPT's "Assistant" feature: OpenAI and Agents are one piece of hardware apart
Gemini 2.0 has already launched a series of functional implementations on AI Studio that can be expanded after obtaining system permissions; Claude models have achieved interaction with various applications through MCP; now it is OpenAI's Agent.
The reason everyone is doing this so consistently is based on the premise of more powerful "Agent" instructions generated by their own models, as well as stronger reasoning, searching, and coding capabilities during the Agent's workflow.
However, in a purely software environment, the functions that can be achieved are quite limited for most people—that belongs to the world of restless programmers.
In my view, so-called AI hardware is not just about breaking through the "cage," but more about lowering the barrier to entry and acquiring enough users. This won't happen immediately, but it is very close.