What Do You Do While AI Codes? I Make Mine Argue With Itself.

A developer built AdversarialDebate, an open-source multi-model code review system in which two LLMs analyze a pull request in parallel with no shared context, commit independent verdicts, and then debate under a structural constraint that prevents either model from anchoring on the other's output. Testing across 411 debates on 70 public-repo pull requests produced an 88.7% binary match against a corrected 2,333-row dataset at a total cost of $0.42 for 360 reviewer runs, and the developer reports the best pairing was GPT-4o-mini with Mistral Small 3.2, arguing that diversity of training objective predicts debate quality better than raw model capability.

Be honest: what do you actually do while the agent types? I used to just watch. Not read, watch . Scroll the diff as it streamed in, nod at code I hadn't fully parsed, and tell myself I'd review it properly at the end. Sometimes I did. Often I didn't. The question "what do you do while AI codes?" keeps showing up in the dev.to feed, and the comments make it clear I'm not alone. The struggle to stop mid-task and just wait is real. So I tried to fill the wait with something useful. I built a system where a second model tries to break the first model's work. Then I learned the thing that changed how I think about AI review: 89% of what I built was theater. I pointed two LLMs at the same pull request and asked them to argue. It worked beautifully: confident claims, point-by-point rebuttals, a clean verdict at the end. Then I read the raw logs. The second model wasn't analyzing anything. It was replaying pre-generated text. Both models were exchanging messages, no positions were changing, no new evidence was being cited. The transcript sounded like a debate. It was a recording of a debate. If I hadn't looked at the raw output, I would have shipped it and called it a working system. That's the part that still bothers me. Here's what most multi-model review workflows actually do: It isn't. It's a validation . The moment Model B's context contains Model A's verdict, B is anchored. It's doing social-pressure resistance, not analysis. Humans fail this constantly. That's why double-blind review exists. We invented that safeguard decades ago and keep leaving it out of AI pipelines. The fix is mechanical: Model B cannot see Model A's output until B has fully committed its own position. Not prompted into independence. Structurally prevented from anchoring. That one constraint is the entire project: AdversarialDebate https://github.com/deghosal-2026/adversarial-debate . Two models analyze in parallel with zero shared context. Each commits a verdict with structured claims and evidence. Then a bounded debate opens: every objection must reference a specific counter-claim, every claim must cite specific text from the artifact. When they can't resolve, the output is a structured disagreement report , both positions preserved and unresolved points documented. That last part matters. Most multi-agent systems are engineered toward consensus. Disagreement is treated as a failure to be smoothed over. But if two independent reviewers reach different conclusions with different evidence, that tension is the signal . Collapsing it into one confident verdict throws away the exact thing you built two reviewers to find. Theater rate across 217 debates in v0.2.0: 0%. I ran the first real field test on 70 actual pull requests from public repos , 411 debates. The core question: do the claims correspond to what actually went wrong in those PRs? By v0.2.0: 150 artifacts across four domains, 88.7% binary match against the corrected 2,333-row dataset, near-zero fabrications. Total cost: $0.42 for 360 reviewer runs. Less than lunch. Compute was never the bottleneck. Prompt engineering and ground-truth measurement were. The best pairing wasn't the two smartest models. It was GPT-4o-mini + Mistral Small 3.2 , a small OpenAI model and a small European one. GPT + Gemini produced the worst debates: lots of rounds, almost no concessions, no resolution. Two RLHF-heavy models from big US labs agree fast because they were trained to prefer harmony. GPT-mini + Mistral produced genuine disputes, real concessions, and eventual convergence. My working theory: it's not raw capability that predicts debate quality, it's diversity of training objective. A later run refined this: the effect may be as much about non-Mistral models rubber-stamping each other, so the practical rule became "always include Mistral." I hold the theory loosely. I have not proven this rigorously. It matches everything in my data, and it's why my model-selection advice starts with "pick from different labs" before anything about benchmark scores. 88.7% sounds good. It is good, compared to single-pass review. But there's a problem: False negatives are invisible. If both models miss a real issue and converge on "looks fine," you get a clean verdict and no indication anything is wrong until it surfaces downstream. The 11.3% that's partial or wrong clusters in the narrative domains: incident reports, change proposals, where ground truth is fuzzier. Generic prompts that work on code don't port cleanly to a postmortem. Requiring two independent models to both miss an issue is a much higher bar than single-pass review. It is not a guarantee, and I won't pretend it is. What do I do while AI codes? I don't watch anymore. I let a second model try to tear the first one's work apart while I do something that actually needs a human: reading the failure modes, writing the eval, deciding what "good" means. It's not passive. It's a different kind of attention. And honestly, it's less lonely than watching a cursor. So what do you actually do while the model types? Do you review live, review after, or have you found a way to use the wait?