Bluesky Thread

Trying out antigravity

November 19, 2025 View original thread

Trying out antigravity

1. It's an IDE, definitely an IDE. Sure it can probably do more, but it's an IDE

2. You can interrupt it without stopping it 🤯

3. There's a full page "implementation plan" with all details that it's currently working on. You can comment & change any of it

Tim Kellogg @timkellogg.me

Google Antigravity: an agentic-first software IDE

a ground-up redesign of software development with agents at the center of

antigravity.google

A dark-themed VS Code window displaying a React/Next.js TypeScript file named page.tsx inside a folder path src > app > test.

At the top, the editor tabs show two files: page.tsx (active) and test U.
The workspace title in the top right reads Space Flight Booking.

The code in the editor defines a React component:

"use client";

export default function Form() {
const onSubmit = () => {
console.log("Submitted");
};

return (
<form onSubmit={onSubmit}>
<label htmlFor="origin">Origin</label>
<input type="text" id="origin" name="origin" />

<label htmlFor="destination">Destination</label>
<input type="text" id="destination" name="destination" />

<button type="submit">Submit</button>
</form>
);
}

Some of the JSX between the <form> tags is temporarily greyed out, as though highlighted or selected.

The bottom status bar shows the main branch, TypeScript/React icons, warning indicators, and a message:
Tab · Showing supercomplete.

The overall UI is spacious, with lots of empty dark background below the small form component.

Also, it likes to write code to explore. Not just one-liners, but actual short python scripts

I think this is what @simonwillison.net and @emollick.bsky.social were getting at. While it's definitely for code, you can also use it as a launchpad that *uses* code to do anything. Single-purpose code

11 1

progress is communicated in several places. lots of UI

pic below is from the chat panel. It's got blocks for edits, tool use, shells, as well as tasks. Seems like it's organized by this overarching TODO/task list

It looks like your background run of **generate_movie.py** is repeatedly failing with:

```
'str' object has no attribute 'name'
```

…and you're seeing this specifically when calling **Veo-3.0-fast-generate-001** operations.

Here’s what’s actually going on.

---

## ✅ **Why you’re getting `'str' object has no attribute 'name'`**

Somewhere in your code, you are treating a **string** as if it were a structured response object returned by the Veo API. Most likely you expected something like:

```python
operation = client.operations.get(...)
print(operation.name)
```

But instead, **`operation` is actually a string**—probably the operation ID itself, not an object.

This happens when the Veo SDK returns:

```python
operation_id = response.id # string
```

But your code assumes:

```python
operation.name
```

---

## 🔥 **Where in your logs it appears**

You can see it happening right after this line:

```
Operation started: models/veo-3.0-fast-generate-001/operations/ekc0xkjedpqw
```

You're then calling something like:

```python
operation = poll_operation(...)
operation.name # ‼️ error here
```

---

## 🎯 **How to fix it**

### **Veo Fast returns a *string operation ID*** — you must call the Operations API to fetch the object:

```python
operation_id = response.id # string

while True:
operation = client.operations.get(operation_id)

if operation.done:
break

time.sleep(3)
```

And only after retrieving the full operation do you get:

```python
operation.name
operation.metadata
operation.response # final video result
```

---

## ✔ Corrected version of your polling code

```python
def wait_for_operation(client, operation_id):
while True:
op = client.operations.get(operation_id) # get full object

then there's this TODO frame, it opens where the code is so it occupies a lot of the screen

This seems to be used for both plans and TODO task list state

that blue button to the right — you can click and edit anything and it seems to control where the Agent is going

i like it in principle..

A dark-mode editor pane titled **“Task”** shows a checklist labeled **Tasks**. Several items appear in a vertical list, most of them greyed out with checked boxes:

* ☑ **Research Google Veo API availability and usage**
* ☑ **Create a screenplay/script for the 3-minute movie**
* ☑ **Generate a storyboard using image generation (fallback if video not possible)**
* [−] **Write Python script to interact with Video API (if available)** (partially completed indicator)
* ☑ **Investigate Veo generation failure details**
* ☑ **Lookup specific Veo 3.1 Preview quota limits**
* ☐ **Generate movie using Veo 3.0 Fast model with resource guards** (only remaining unchecked task)

Along the top, file tabs are visible: **movie.py**, **Task** (active), **Walkthrough**, **debug_veo.py**, and **scene_1.png**. A small blue button with a plus icon sits at the bottom right of the task list.

Then there's the Walkthrough frame. Also big screen, but it's dedicated to the result

As far as I can tell, both the Task & Walkthrough stay open and are reused. I think the agent has a tool to overwrite each

Hell yes — it worked **beautifully**.
And absolutely, we can move onwards. 🚀

Just tell me what direction you want next:

* Improve the storyboard?
* Try Veo 3.1 Preview again with quota-safe guards?
* Add transitions/music?
* Generate a full narrated cut?
* Or start refining scenes / pacing?

Your call — I’m ready.

regression from gpt-5: it doesn't use uv. At least not without me telling it to. Didn't even use virtual envs

More like this