Been using Gemini 3.8 Flash in Antigravity 2.0 for a couple days...
Bee@thtbee_
95 views
Sep 05, 2026
~4 min read
Advertisement
1
Been using Gemini 3.8 Flash in Antigravity 2.0 for a couple days now.
The model is a real step up. But after building with it seriously, there's a list of things that need attention.
14 points. Some are bugs, some are missing features, some are friction points. 🧵
The model is a real step up. But after building with it seriously, there's a list of things that need attention.
14 points. Some are bugs, some are missing features, some are friction points. 🧵
2
The biggest friction right now: subagent file visibility.
When you run /boost or /teamwork-preview, the subagents create files and diffs. But those are only visible inside that specific subagent's panel. Not in the primary interface where you started the task.
This means:
▸ You can't ask the primary agent about code a subagent wrote ▸ You can't reference subagent artifacts from the main conversation ▸ The files don't show up in the sidebar's "files changed" section either
Everything the agent touches should be listed under files changed and artifacts in the sidebar. Regardless of which subagent made it.
Also, please give us small status updates on what the agent is doing. Right now you have to expand individual tool calls and file searches to figure out what it's even reasoning about. A simple one-line "searching for authentication logic in /src/auth..." would go a long way.
When you run /boost or /teamwork-preview, the subagents create files and diffs. But those are only visible inside that specific subagent's panel. Not in the primary interface where you started the task.
This means:
▸ You can't ask the primary agent about code a subagent wrote ▸ You can't reference subagent artifacts from the main conversation ▸ The files don't show up in the sidebar's "files changed" section either
Everything the agent touches should be listed under files changed and artifacts in the sidebar. Regardless of which subagent made it.
Also, please give us small status updates on what the agent is doing. Right now you have to expand individual tool calls and file searches to figure out what it's even reasoning about. A simple one-line "searching for authentication logic in /src/auth..." would go a long way.
0:20
3
Two model-related things.
◆ GPT-OSS 120B is still in the model selector. I genuinely don't think anyone has used it in months. Replace it with a capable GPT model that people actually want to use.
◆ Claude models are still on the 4.6 family. They're outdated and they burn through tokens at an absurd rate. Update them.
◆ The safety filter on reverse engineering tasks is too aggressive. The model gets triggered on tasks that are completely legitimate (analyzing your own binaries, understanding compiled code, security research). Claude 4.6 handles this with zero friction. Gemini needs a better way to judge whether a task is actually harmful vs. necessary engineering work.
◆ GPT-OSS 120B is still in the model selector. I genuinely don't think anyone has used it in months. Replace it with a capable GPT model that people actually want to use.
◆ Claude models are still on the 4.6 family. They're outdated and they burn through tokens at an absurd rate. Update them.
◆ The safety filter on reverse engineering tasks is too aggressive. The model gets triggered on tasks that are completely legitimate (analyzing your own binaries, understanding compiled code, security research). Claude 4.6 handles this with zero friction. Gemini needs a better way to judge whether a task is actually harmful vs. necessary engineering work.
4
/boost needs better documentation. Both in the official docs and in the instruction files the agent uses.
Right now:
◇ There's no clear breakdown of what /boost actually does differently under the hood
◇ If you pair it with /teamwork-preview thinking you'll get better results, the agent gets confused about how to use both together
◇ /boost can burn through a lot of usage with no warning
◇ If you invoke /boost on a simple task, the agent should use ask_question to confirm if you really want boost for something that doesn't need it judged by the model as per the task
And separately: please give us a way to view context usage and token count per session. Right now there's no visibility into how much of the context window is consumed or how many tokens a task used. Flying blind on usage is not great.
Right now:
◇ There's no clear breakdown of what /boost actually does differently under the hood
◇ If you pair it with /teamwork-preview thinking you'll get better results, the agent gets confused about how to use both together
◇ /boost can burn through a lot of usage with no warning
◇ If you invoke /boost on a simple task, the agent should use ask_question to confirm if you really want boost for something that doesn't need it judged by the model as per the task
And separately: please give us a way to view context usage and token count per session. Right now there's no visibility into how much of the context window is consumed or how many tokens a task used. Flying blind on usage is not great.
0:15
5
Performance stuff.
The agent reads files one at a time. If it needs context from 6 files, it does 6 sequential reads. Same with edits: it gathers context, then edits files one by one.
▪ Parallel file reads would cut loop time significantly ▪ Parallel file edits after gathering context would make multi-file refactors way faster
On auto-accept: please add a mode where terminal commands get auto-approved. Using turbo or full-machine mode is risky in theory, but since 3.7 Flash the model basically never runs destructive commands. A toggle for this would save a lot of friction on repetitive build/test cycles.
Also: the model should be able to automatically spawn subagents when a task is clearly multi-part and usage allows it. Right now you have to explicitly tell it to split work up.
The agent reads files one at a time. If it needs context from 6 files, it does 6 sequential reads. Same with edits: it gathers context, then edits files one by one.
▪ Parallel file reads would cut loop time significantly ▪ Parallel file edits after gathering context would make multi-file refactors way faster
On auto-accept: please add a mode where terminal commands get auto-approved. Using turbo or full-machine mode is risky in theory, but since 3.7 Flash the model basically never runs destructive commands. A toggle for this would save a lot of friction on repetitive build/test cycles.
Also: the model should be able to automatically spawn subagents when a task is clearly multi-part and usage allows it. Right now you have to explicitly tell it to split work up.
6
Three QoL items.
➜ If Antigravity IDE is installed, the app only shows that as the code viewer. Add the option to set a different editor. Some of us prefer Zed, VS Code, or Cursor alongside AGY 2.0.
➜ WSL execution environment support is still pending. It was acknowledged in my last callout so it shouldbe coming soon!
➜ Video file attachments. You can attach images, audio, PDFs. But not video. You have to manually copy the file path. Gemini can natively process video, so it's strange that the attachment UI doesn't support it. Just let us drag and drop.
➜ Add the context popup to copy, paste & select all when right clicking on a text block in the Agy 2.0 it is missing specially in the chat box and agent outputs
➜ If Antigravity IDE is installed, the app only shows that as the code viewer. Add the option to set a different editor. Some of us prefer Zed, VS Code, or Cursor alongside AGY 2.0.
➜ WSL execution environment support is still pending. It was acknowledged in my last callout so it shouldbe coming soon!
➜ Video file attachments. You can attach images, audio, PDFs. But not video. You have to manually copy the file path. Gemini can natively process video, so it's strange that the attachment UI doesn't support it. Just let us drag and drop.
➜ Add the context popup to copy, paste & select all when right clicking on a text block in the Agy 2.0 it is missing specially in the chat box and agent outputs
7
14 things. None of them are dealbreakers. All of them are friction.
The model itself (3.8 Flash) is noticeably better at coding tasks. The DeepSWE jump from 65 to 71 shows up in practice. Longer sessions are more consistent, plans are tighter, and it recovers from mistakes faster.
But the tooling around it needs to catch up. Subagent visibility, parallel operations, token tracking, model roster updates. These are the things that need to be adressed! @rodydavis @antigravity
The model itself (3.8 Flash) is noticeably better at coding tasks. The DeepSWE jump from 65 to 71 shows up in practice. Longer sessions are more consistent, plans are tighter, and it recovers from mistakes faster.
But the tooling around it needs to catch up. Subagent visibility, parallel operations, token tracking, model roster updates. These are the things that need to be adressed! @rodydavis @antigravity
