The AI Teardown· August 4, 2026

AI Code Review: Why Review Direction Beats Model Size

Give Claude a stack of Codex’s code to review and the pass rate climbs from 71.6% to 89.7%. Flip the roles and Codex reviewing Claude’s drafts drags a 91.4% baseline down to 82.8%.

That asymmetry, pulled from a paper the day’s top thread corrected in real time, settles a question most agent stacks answered backwards. Cross-checking agents is not a democracy. It is a hierarchy, and the direction of review decides whether you gain accuracy or burn money destroying it.

The org chart, not the ensemble

The mechanism is blunt: a stronger model reviewing a weaker one lifts quality, and a weaker model reviewing a stronger one destroys it. Claude working alone scored 91.4%, the best of any configuration, and reviewing its own work moved that number nowhere.

The consequence lands directly on a popular cost-saving pattern. Every cheap-model-as-reviewer scheme built to save tokens is, whenever the reviewer is the weaker model, actively degrading output while charging for the privilege. The token saving is real; so is the quality bill, and only one of them appears on the invoice.

What survives contact with the finding is a loop pointed uphill: one builder has Claude draft a plan, a stronger external model tear it apart, and Claude replan until they agree, at which point it “just basically spits out the code.”

Duration is the new capability

If review direction is the mechanism, sustained looping is the workhorse, and the week’s most credible autonomy claim arrived with receipts attached. An audit loop that works one verified task per iteration ran against 17 open-source projects and drove 16 to convergence across eight languages, with an adversarial sub-agent signing off each pass.

The receipts are checkable by anyone: a security flag in bat that did nothing when output was piped, and a Content-Length bug in fasthttp that turned a rejectable number into a wrong one, both fixed and merged upstream.

The sharpest comment in the thread restated the org-chart law from the failure side: a verifier that shares the generator’s blind spots approves fixes that miss the intent. Alibaba’s ten-day unattended Qwen run wears the same lesson as a headline. The loop that grinds and self-corrects beats the model that nails it once.

The cheapest defence is an order swap

The most reusable trick of the week costs almost nothing. A three-tier photo culler running Haiku, Sonnet, then Opus across a 25,000-image library for about $25 has Opus judge each head-to-head pair twice with the order swapped, counting disagreement as a tie.

Position bias turned out to be arithmetic rather than theory: about one in eight pairs flipped on the swap and got correctly nulled. Two judgments with the order swapped, and the LLM-as-judge failure that quietly corrupts evals stops corrupting yours.

It pairs with the week’s best ops note: a frozen set of labelled examples run before any prompt or model change, so the pipeline cannot slowly learn a new taste without anyone noticing.

Suggest is cheap; commit is not

All of it stays academic until an agent commits an action instead of suggesting one, and someone went looking for evidence. Of 396 European fintechs checked for proof their AI agents actually act, writing to a system of record, moving money, closing a case, only 143, about 36%, could show it. The other 64% are copilots wearing the word agent.

What nobody in that 36% publishes is what happened when the committing kind got it wrong, and that silence is the actual risk. The week’s most concrete failure spec, from a standing-team agent server, names the danger precisely: stale in-process context outranking a newer decision record, an idle agent woken with old assumptions executing queued work the organisation already superseded.

Suggest-and-a-human-commits wastes time. Commit-and-a-human-reviews-later ships a wrong irreversible action with things already built on top of it, and the difference between those two sentences is the entire agent-safety debate.

The quick cuts

Eight AI memory systems, 2,176 tasks, and the winner was not a product: a plain markdown wiki the agent curates itself scored 98.5 and beat every paid system. One commenter called it: “the agent just wants a notepad.”

Five web-search APIs wired as MCP tools and scored on identical questions produced the week’s bill: Keiro at $0.25 per 1,000 calls matched the 8.5 quality of Exa and Tavily, which charge $7 and $8. A 28x price gap for the same scored quality, minus a naming clash that hijacks the agent’s built-in search.

The sentiment story: 338 upvotes agreed that Opus 5 argues too much, with users pinning Opus 4.8 instead. The OP’s plea, verbatim: “Bro, just do the thing I asked for…” The nuance worth holding is that the same community recommends Opus 5 for generative design: a personality regression for iteration, a tool for one-shot creation, and a routing decision either way.

By the numbers

The close

The builders pointing review uphill and swapping judge order this quarter get the paper’s 18-point lift for the cost of reading it. The stacks saving tokens with a cheap reviewer are paying a premium model to be corrected downward by its understudy, and the invoice will never once mention it.