Most AI agent ideas are lame; they’re essentially a glorified Zapier (I don’t hate Zapier; it has its place). I have been struggling to understand what a good AI agent should look like.

I have been using Cursor for a few weeks, and here is why it’s a good AI Agent.

* It is a complete experience: Cursor did not integrate into VS Code as a plugin; they just forked the damn thing—owning the coding experience. Cursor does not need permission from VS Code; it sees everything, and hence it’s better.

* It’s actually doing work: It’s not providing information to a human so they can do it better. It actually does the work a human wants to be done. Coding agents write code; support agents should resolve issues.

* Narrow the use case: Things that don’t work well in coding agents are edges where your code speaks with other systems. This takes lots of tries and may eventually need human effort.

* It can get complex jobs done with some back and forth: It’s right on the first try about 20% of the time, and correct eventually 90% of the time with some back-and-forth.

* Lets humans focus on the “what” (not the “how”): When working with an AI agent, the engineer can focus on what needs to be done, while the agent handles how it needs to be done.

* Does work better than the bottom half: Code written by AI is often better than that written by the bottom 40% of human engineers.

* I do not see AI replacing engineers, but I see human engineers becoming even more productive using AI agents. This applies to every field impacted by AI.

It’s clear that in a few years, coding without an agent will be like not using an IDE.

HN, where have AI agents slayed your toughest workflows—and where have they flat-out failed? Share your war stories and wins!

Btw this is my startup https://www.twig.so and no, we don't do this yet. We are crawling towards it.