DEV Community

Take Claude Code Above Average: Let Antigravity Review It

Daniel Gwerzman on November 09, 2026

In November 2026, I spent a day building a small tool with Claude Code: a bridge that lets a second AI agent review Claude's work, run the tests in...
Collapse
 
innokentyb profile image
Kent Bodrov •

The pushback step seems as important as having a second reviewer. How do you measure whether it improves the outcome rather than just producing more findings? A small seeded-defect set plus checks for false positives and regressions introduced by accepted fixes could make the review loop testable. I would also keep those tests separate from the examples used to tune the prompts.

Collapse
 
paullovelace profile image
Paul Lovelace •

The uncommitted-work finding is the one that made me wince — a reviewer with shell access and one stray git reset --hard from wiping everything. That's not a review bug, that's a missing seatbelt. I'm curious though: now that every review goes to Gemini 4 Argon, have you hit a case where the reviewer was confidently wrong and Claude just obeyed it? The second model can be wrong too, just differently.

Collapse
 
kulaone profile image
Daniel Gwerzman Google Developer Experts •

In some cases Argon got it wrong, but then Claude refused it.
It is not clean from mistakes, but I'm getting better results.

Collapse
 
maxritter profile image
Max Ritter •

Your point that the author and its second pass share the same blind spots matches what I see. I run review with an agent that did not write the change, and I give it the numbered done list we agreed before coding, so its findings tie back to what was asked. Decisions from those reviews go into an MCP memory server, so the next session starts with them. I built QualityLayer for this: github.com/maxritter/qualitylayer

Collapse
 
gapmine profile image
Balance •

Really liked the first finding, because it turns review safety into a boundary problem, not just a backup problem. A backup helps recover, but a disposable worktree or snapshot gives the reviewer room to run tests without touching the author's uncommitted tree. I'd want the final diff checked before anything gets merged back.

Collapse
 
raju_dandigam profile image
Raju Dandigam Google Developer Experts •

I like the rule that a reviewer's finding is advice to verify, not a command to apply. When Claude pushes back, do you persist the finding, counter-evidence, test result, and final decision with the PR? That would make a later regression understandable without replaying the whole conversation.

Collapse
 
arthur2122 profile image
Arthur Figueiredo • • Edited

I liked the point that the author can't be the reviewer. My flow is similar, but inverted: Claude plans the specs, a separate agent writes the tests first, agy (Antigravity) writes the code, and Claude comes back to check it. Take a look at my flow: ArthurFigg/claude-dev-flow (the repo is in Portuguese because I'm Brazilian). Using two AIs that think differently improves things a lot (I'd even like to add more AIs that think differently, but it gets more laborious), and it should give you some good insights.

Collapse
 
mudassirworks profile image
Mudassir Khan •

the two findings your reviewer caught, two sessions starting their own process and no backup before it could run shell commands, are both the boring operational kind an author never sees in its own design. i'd add that the reviewer needs a read only default, because the first time it runs a destructive command on your uncommitted work the second opinion becomes the incident. does the bridge run the reviewer in a worktree or a copy, or straight in the project directory?

Collapse
 
kulaone profile image
Daniel Gwerzman Google Developer Experts •

It depends on your choice. You can prompt it however you prefer.

Collapse
 
parsam profile image
Parsa •

The core issue with having an agent review its own code or even a diff from another model is that it only judges changed lines. The expensive failures usually sit in code the PR never touched, like an upstream caller assuming an old return shape or a shared resource not backed up before running shell commands. Line review has no visibility into how a change alters the wider system state.

Collapse
 
kulaone profile image
Daniel Gwerzman Google Developer Experts •

This is very easy to solve with the right skill. Just ask it to check the entire repo.

Collapse
 
dpietro profile image
Denise Pietro •

Great write-up. The “author can’t be the reviewer” framing is the part that stuck with me, and the two-sided obedience trap is a sharp way to put it: a model that agrees with its own work and a model that blindly applies its reviewer’s suggestions are failing in the same way.

The git apply --relative example was the most convincing moment for me. Right bug, wrong fix, and the pushback only worked because Claude tested the suggestion in a scratch repo instead of just applying it. That’s a good argument for treating findings as advice to verify rather than instructions.

I also like requiring a named source on every assumption tag. The detail that the first version accepted guesses as facts until you insisted on a source shows how small prompt details change whether the guard works at all.

One question: have you seen the reviewer’s findings drift when the author’s brief is too persuasive? If the lead agent frames the work confidently in Stage 1, I’d expect some of that confidence to leak into the review despite the clean slate. Do you do anything to keep the brief neutral, or does the outside-expert framing handle it?

Thanks also for being upfront about the limits, like shell commands not being restrictable and the risk of planted skills in untrusted repos. That honesty makes the rest more credible.