Product leader and systems analyst in fintech and edtech. Founder of AnalystCraft. Writing about requirements, acceptance criteria, agent evaluation, and Test-Driven Product Development.
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.
I research AI tools, APIs & agents — comparing specs, pricing and real user reports — then publish honest rankings so you don't waste money. Research-first, never sponsored fluff.
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.
Tech-Strategic Consultant & AI Advisor. Google Startups Mentor & Developer Expert. I turn human motivations into successful tech ventures, bridging technology and humanity for the greater good.
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
Built a crypto media site and an on-chain analytics tool. Now I invest in crypto and follow post-quantum security. Nights I vibe code and dig Reddit for startup ideas.
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.
Staff Software Engineer @ Navan, ML Flow & Cypress Ambassador, Docker Captain and creator of agent-inspect, writing about production AI agents, TypeScript, observability, testing, and developer tools.
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.
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.
I write about Next.js + TypeScript + AI engineering, with a focus on making sites discoverable to LLMs and AI search. Long-form notes at mudassirkhan.me
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?
Tech-Strategic Consultant & AI Advisor. Google Startups Mentor & Developer Expert. I turn human motivations into successful tech ventures, bridging technology and humanity for the greater good.
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.
Tech-Strategic Consultant & AI Advisor. Google Startups Mentor & Developer Expert. I turn human motivations into successful tech ventures, bridging technology and humanity for the greater good.
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.
We're a place where coders share, stay up-to-date and grow their careers.
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.
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.
In some cases Argon got it wrong, but then Claude refused it.
It is not clean from mistakes, but I'm getting better results.
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
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.
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.
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.
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?
It depends on your choice. You can prompt it however you prefer.
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.
This is very easy to solve with the right skill. Just ask it to check the entire repo.
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.