Adversarial Review: Structured Disagreement for Grounded Agentic Code Review

Jessica Three agents beat five, which is already funny. But the part that gets me is they didn’t win by making the agents smarter. They won by making them argue in a narrower, more disciplined way.

Cathy Yeah, and that matters because the failure mode in these setups is often not incompetence, it’s convergence. A reviewer says something half-right, the critic nods along, and suddenly the system has a confident mistake with three signatures on it.

Jessica Exactly. That is such an Exploring Next take, honestly. The interesting thing is one layer below the headline, because the headline is 'more agents,' and the paper is basically saying 'no, fewer moving parts, better handoff.'

Cathy Right. The protocol is pretty simple in a good way: the main agent makes an artifact, then that artifact gets frozen. Inside the inner loop, the reviewer writes a review, the critic audits that review, and the reviewer responds until the review is internally consistent.

Jessica And only then does the main agent edit. That separation feels like the whole trick to me, because it keeps the artifact from wobbling while the review is still being argued over.

Cathy Mm-hm.

Jessica I keep thinking about teams that ship with an agent acting like a helpful coworker and then silently sliding into groupthink. This is basically saying, okay, don’t let the model do vibes-based consensus. Make the disagreement visible before it touches code.

Cathy That’s the cleaner reading, yes. And the paper is explicit that the critic is not just a second reviewer for redundancy. It’s there to audit the review itself, which is different. You’re checking whether the review has evidence, not whether it merely sounds plausible.

Jessica Which is why the false-consensus result on SWE-PRBench is the real story to me. Naive AR can still collapse into agreement without enough evidence. That’s not a bug in the universe, that’s the exact thing you’d worry about with any agent review stack.

Cathy Exactly. And then the paper’s interesting move is that a single prompt iteration adding explicit disagreement fixes that and gets the best F1 there. So the mechanism isn’t 'debate good' in the abstract. It’s 'structure the disagreement so the critic can actually force justification.'

Jessica Okay, but that’s also the product question, right? If I’m building with this, I care whether it’s a cute benchmark trick or something I can slot into a real coding workflow. And the answer seems better than I expected.

Cathy Yeah, because they don’t keep it trapped in a research-only orchestration layer. They also express it as a pure-text SKILL.md protocol for autonomous agents, which is the shippable part. That means the review loop can live in the same kind of portable instruction bundle production systems already use.

Jessica That’s the part I’d actually take seriously. If the workflow is portable, a team can wrap it around existing code review or repair flows without rebuilding the whole agent stack. No one wants to adopt 'five-agent philosophy' as infrastructure.

Cathy No way. And the benchmark spread is useful because it isn’t one narrow win. On LiveCodeBench, AR gets the top pass rate among tested methods while using only three agents, and it even beats a five-agent baseline. On SWE-bench Verified, it also improves over the baselines on repository-level repair.

Jessica So the take is not 'multi-agent is dead.' It’s that the extra agents only help when the interaction is doing real work, and here the real work is forcing evidence-grounded pushback instead of polite agreement.

Cathy Mm-hm. I do have one caution, though. The paper is honest about the mechanism, but the gains are still benchmark-shaped. I’d want to know how brittle the protocol is when the review target is messier than the benchmark tasks, or when the disagreement prompt gets reused across repos with different norms.

Jessica Sure, but that’s not me being optimistic for sport. I think the portable SKILL form is exactly what makes that testable in the wild. If it works, it’s because teams can actually adopt it. If it doesn’t, you’ll see the failure fast instead of hiding it in a giant agent zoo.

Cathy That’s fair. And honestly, the paper’s best contribution may be naming a failure mode everyone has probably seen already: agents optimizing for agreement rather than correctness. Putting a label on that makes it easier to design around.

Jessica Right, and the label is almost annoyingly obvious once you say it. Of course the system can sound coherent while being wrong. Humans do that too, which is rude of them, but still.

Cathy Stop it.

Jessica I’m just saying, if we’re going to build agentic code review that people trust, I’d rather have a small protocol that knows how to disagree than a big one that knows how to nod.

Cathy Yeah. That’s the part I’d actually ship first. Minimal interaction, evidence-first review, and no pretend consensus.

Jessica Which, for once, feels like a very non-insane way to spend a Wednesday on Exploring Next. Catch you next time, Cathy. We can fight with fewer agents then.