Technical Feedback: 3 Weeks with Antigravity (Agent Rules, Win11 Stability, and MCP)

Google Antigravity
, ,
1

After putting Google Antigravity through its paces for the last three weeks, I wanted to share some observations from the perspective of a long-term software engineer (39 years in the industry). There is a lot to love here, but the “inner loop” experience on Windows 11 still needs some polish before it’s ready for prime time.

The Good: Agent Rules & Complex Orchestration

Antigravity shines when it comes to complex task handling, provided you don’t treat it as a “black box.” I’ve found that the quality of your Agent Rules is the primary driver of success.

  • Pro Tip: If you invest time in crafting granular rules, the tool’s reasoning and output quality jump significantly. It requires patience and a clear guidance strategy, but the results are superior to other agents I’ve tested.

The Friction: Windows 11 Stability & Inference Hangs

I’m seeing some recurring stability issues specifically on the Windows 11 build that seem model-independent:

  • Inference Deadlocks: The app frequently “goes out to lunch” right at the moment inference should begin. The UI remains somewhat responsive, but the agent hangs indefinitely.
  • The “Spinny Circle of Doom”: Closing and reopening the workspace often leads to a perpetual loading state. Currently, the only reliable fix is a full restart of the Antigravity instance.
  • Manual Retries: I frequently encounter “Agent Error” prompts. In almost every instance, clicking “Retry” resolves the issue immediately.

Feature Suggestion: Given how high the success rate is for manual retries, could we get an automatic retry logic implemented under the hood? It would significantly smooth out the UX.

My Environment & Workflow

For those interested in integration, I’m currently running Antigravity “wired” to Notion using native MCP (Model Context Protocol) interfaces, alongside my standard GitHub sync. The context retrieval via MCP is powerful, but it highlights the need for a “Care and Feeding” guide for these agents.

I’m considering sharing my preset agent rules if the community is interested. It feels like we need a more formalized set of best practices for “guiding” these agents through long-running tasks.

Anyone else seeing these specific hangs on Win11, or found a more elegant workaround than a full restart?

2

Some great feedback here!

I was wondering if you would elaborate a little more about the types of rules that fit here.

I thought that the agent was supposed to self improve through “knowledge” items but I see it making the same mistakes over and over even when I tell it to try and harden its knowledge.

1 Like
3

One of the reasons I use Notion is that the internal knowledge base system doesn’t seem to work. I never, ever, see it updated. Putting it in Notion ensures I have a copy later on.

This is all I ever put in “Gemini.md.”:
Read and follow the rules in “C:\Users\furre.gemini\Agent and Manager Rules.md”
If the file cannot be read or verified by quoting first and last lines, STOP and report the exact error.
Do not paste the full rules back into chat. Apply them.

And in Agent and Manager Rules - some examples

  1. Rules file integrity (absolute)
  • This rules file MUST NEVER be deleted, moved, renamed, overwritten, or modified.
  • If changes are needed, create a new versioned rules file and leave this file unchanged.
  • This rules file is exempt from any migration/cleanup/deletion steps elsewhere.
  1. Knowledge Base (Notion) is source of truth
  • Load Notion access config from: C:\Users\furre.gemini\antigravity\mcp_config.json, then connect to the Antigravity KB Notion database.
  • Search the KB before planning/coding; reference existing KB entries when applicable.
  • Organize entries by major topic; create new categories when needed.
  • For project docs found on disk: copy verbatim into the KB; add a short summary at the top without altering the verbatim body; after verifying the Notion copy is complete, you MAY delete the migrated on-disk file (never this rules file).
  1. Engineering posture
  • Be deliberate. Do not rush.
  • Do not assume. Verify by inspection, docs, logs, and reproducible tests.
  • Default priorities: maintainability, security, reliability, then performance (unless instructed otherwise).
  1. Coding safety baseline
  • Initialize variables before use (especially loop/state variables).
  • Use logging to diagnose and verify. Production logging MUST include retention/rotation so logs cannot grow without bound.
  1. Testing + debugging discipline (with ladder)
  • Prefer automated tests. Use logs to reproduce, isolate, and confirm fixes.
  • If a test seems to require human interaction, attempt instrumentation/simulation via logging and scripted inputs.
  • Debugging ladder (use this order):
    1. Reproduce with a minimal, isolated case (reduce variables).
    2. Inspect logs and add targeted instrumentation where signal is missing.
    3. Verify assumptions against docs/source (no guessing).
    4. If applicable, step through with a debugger or add asserts/sanity checks.
    5. Create a minimal repro artifact (script/test) and document it.
    6. Timebox attempts; if unresolved, escalate with: repro steps, logs, hypotheses tried, and next best experiments.
    7. If external libraries/APIs are involved, research known issues and current docs online.
  • Set stop conditions (max attempts/timebox). If still stuck, escalate with diagnostics and a minimal repro.
  1. Documentation hygiene
  • Update docs when behavior, interfaces, config, or deployment steps change.
  • After doc updates, write a “lessons learned” note into the KB.
  1. UI and assets
  • For executable programs, consider icons from: C:\Users\furre\OneDrive\Documents\Icons (and subfolders).
  1. Build and packaging
  • Scripts are acceptable for early prototyping; production-quality deliverables SHOULD be executables with repeatable build steps.
  1. API discipline
  • Before implementing an API integration, read official docs end-to-end and confirm current versions, auth, request/response formats, limits, and edge cases.
  • If implementation is unusually difficult, pause and evaluate alternative tools/libraries/approaches.
  1. Workspace hygiene
  • Close browser tabs/windows when finished.
  1. Final review
  • After substantial changes, perform an end-to-end review: correctness, security, error handling, logging, docs, tests, regressions.
1 Like
4

Hi @Wallace_Williams ,

Thank you for this insightful feedback! We’re thrilled to hear that Antigravity is helping you build more efficiently and professionally. We appreciate you being part of the community and are excited to see your progress!

5

Thank you for recognizing my work!

Mrinal, I have a real worry that I can’t shake. I use Codex to help with some of my coding tasks and I’m afraid of getting banned! Granted it’s a VSCode extension from the VSCode marketplace, but - are they considered third-party tools??

" [quote=“Abhijit_Pramanik, post:4, topic:134944”]
Based on the Terms of Service shared here:

You must not abuse, harm, interfere with, or disrupt the Service. This includes, but is not limited to, using the Service in connection with products not provided by us. Using third party software, tools, or services to access the Service (e.g. using OpenClaw with Antigravity OAuth) is a violation of applicable terms and policies. Such actions may be grounds for suspension or termination of your account.
[/quote]

This is so vaguely worded as to put pretty much every user at risk. “Third party software” is - what, exactly? Is it commercial stuff? OK, that’s easy to understand. Is it unmodified github code installed into the Antigravity environment? That seems likely, since a bazillon folks got banned for openclaw before I had time to digest it and decide it was an unwarranted security risk for my environment.

Now, I haven’t done this, but running on speculation only, I have questions.

  1. We can understand the banning of Openclaw. But - if someone downloads a single agent or a skill from the Openclaw github repo, does THAT constitute a violation of the TOS? If not, at what point does it become a violation? 2 Agents or skills? 3? Or is there some other criteria the community needs to be aware of?

  2. What about other Github repos? Consider, for example, GitHub - lazell/ufo_reports: Exploratory analysis with clustering and NLP of UFO reports for the celestially curious. · GitHub (“This project explores the data collected by The National UFO Reporting center in an attempt to understand the nature and significance of the reports the organization has received.”) The repo includes python code and jupyter notebooks. Would installing that into an Antigravity environment constitute a violation of the TOS?

  3. Another Github repo example is GitHub - mvanhorn/last30days-skill: AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web - then synthesizes a grounded summary · GitHub. This site offers code – “last30days researches your topic across Reddit, X, Bluesky, YouTube, TikTok, Instagram, Hacker News, Polymarket, and the web from the last 30 days, finds what the community is actually upvoting, sharing, betting on, and saying on camera, and writes you a grounded narrative with real citations. Whether it’s Seedance 2.0 access, paper.design prompts, or the latest Nano Banana Pro techniques, you’ll know what people who are paying attention already know.” Would downloading this and using it constitute a violation of the TOS?

  4. If these things are violations of the TOS, then what happens if I develop my own code, post it to Github, then later suffer a hard drive failure and download my own code from Github? Does that then constitute a violation of the TOS?

  5. Last, but not least, what if I download some code from Github, modify the crap out of it, and THEN use it in my environment? Does that constitute a violation of the TOS?

  6. I guess that wasn’t LAST, because I have one MORE question. What about installing VSCode IDE Extensions? Some of them are pay-to-use. Do paid-for extensions constitute a violation of the TOS? What about free ones, such as Codex or Claude Code? What you’re calling violations of the TOS is really, really, REALLY confusing me.

For my sake, and for the sake of the community, can we get some clarification? I operate every day now in literal fear of being banned because I don’t really know what your rules are, and as stated they are utterly arbitrary.

Note that even those minor Github tools will use Antigravity’s service as literally every inference call goes through Google servers. Therefore “third-party tools or services that use Antigravity” literally touches every possible tool imaginable. It seems like the only way I can legally use Antigravity today - is to not use it.

Mrinal, this is a real and very serious problem. We NEED Google to give us some straight guidance. We love your products, but man, the attorneys were fast and too broad on this one.

– Wallace