The AI Teardown· August 4, 2026

Your Agent’s Reviewer Has to Outrank It

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 it falls: 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, is the most useful thing a builder learns today. Cross-checking agents is not a democracy but a hierarchy, and the direction of review decides whether you gain accuracy or burn it.

The org chart, not the ensemble

The paper behind that correction makes a claim most agent stacks get backwards. A stronger model reviewing a weaker one lifts quality; a weaker model reviewing a stronger one destroys it.

The abstract, quoted verbatim in the comments, reads: “Claude review raises Codex drafts from 71.6% to 89.7%; Codex self review raises them to 84.5%.”

Claude working alone scored the best of anyone at 91.4%, and reviewing its own work left that number untouched.

The consequence is blunt.

Every cheap-model-as-reviewer scheme built to save tokens is, when the reviewer is the weaker model, actively degrading output.

What survives is a loop pointed uphill: one builder has Claude draft a plan, a stronger external model tear it apart, then Claude replan until they agree before it, in his words, “just basically spits out the code.”

Duration is the new capability

If review direction is the mechanism, sustained looping is the workhorse. The most credible autonomy claim today comes with receipts: an audit loop that works one verified task per iteration until a fresh audit is clean and an adversarial sub-agent signs off.

Run against 17 open-source projects, 16 converged across eight languages.

It caught 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 fixes are merged upstream, checkable by anyone.

The sharpest comment is a warning that restates the org-chart finding: when the verifier is the same model as the generator, it shares the generator’s blind spots and approves fixes that miss the intent.

Alibaba’s ten-day unattended Qwen run is the same lesson worn as a headline. The loop that grinds and self-corrects beats the model that nails it once.

The cheapest defense is an order swap

If a same-model reviewer is unreliable, the cheap defense is to stop trusting any single judgment. The single most reusable trick here comes from a three-tier photo culler that runs Haiku, Sonnet, then Opus across a 25,000-image library for about $25.

In the final round, Opus judges each head-to-head pair twice with the order swapped and counts disagreement as a tie.

Position bias is not theoretical: about one in eight pairs flipped on the swap and got correctly nulled out.

Judge twice, swap the order.

That one move catches the LLM-as-judge failure that quietly corrupts evals.

It pairs with the day’s best ops note: a frozen set of labeled examples run before any prompt or model change, so you notice when the pipeline slowly learns a new taste.

Suggest is cheap, commit is not

All of this stays academic until an agent commits an action instead of suggesting one.

Someone checked 396 European fintechs for real evidence their AI agents actually act: write to a system of record, move money, close a case. Only 143, about 36%, could show it. The other 64% are copilots wearing the word agent.

Almost nobody publishes the part that matters: what happened when the committing kind got it wrong.

That silence is the real risk. The most concrete failure spec today comes from a standing-team agent server, where the danger is stale in-process context outranking a newer decision record.

An idle agent woken with old context should refuse queued work built on a decision that has since been superseded.

Suggest-and-a-human-commits merely wastes time; commit-and-a-human-reviews-later ships a wrong irreversible action with things already built on top.

The doctrine the day hands you: point review uphill, loop instead of one-shotting, swap the order before you trust a judge, and never let an agent commit what you cannot recover.

Everything else is a demo.

The Quick Cuts

A builder ran eight AI memory systems through 2,176 tasks. The winner was not a product: a plain markdown wiki the agent curates itself, scoring 98.5 and beating every paid system.

As one comment put it, “markdown wiki winning is the funniest outcome, we build all this complex memory infra and the agent just wants a notepad.”

Five web-search APIs, wired as MCP tools and scored on the same questions. Keiro landed the eye-popping number: $0.25 per 1,000 calls at the same 8.5 quality as Exa and Tavily, which charge $7 and $8.

A roughly 28x price gap for equal scored quality, minus a naming clash that hijacks the agent’s built-in search.

The sentiment story: 338 upvotes agreeing that Opus 5 argues too much, with users pinning Opus 4.8 instead. OP’s plea, verbatim: “Bro, just do the thing I asked for…”

The nuance to hold: the same community recommends Opus 5 for generative design. A personality regression for iteration, a tool for one-shot creation.

By the numbers