You have 1 article left to read this month before you need to register a free LeadDev.com account.
Estimated reading time: 3 minutes
Key takeaways:
- Reviewer hierarchy among AI-coding agents beats having a second opinion at all: the wrong pairing dragged Claude’s 91.4% down to 82.8%.
- Codex fixed 3 of Claude’s failures but broke 13 correct solutions – junior-rewriting-senior, not senior-reviewing-junior.
- A second AI reviewer isn’t free: cost more than doubled, latency nearly tripled.
Putting two AI-coding agents into a writer-reviewer pipeline seems like a logical way to improve code quality: one produces the code, while the other catches its mistakes.
However, that might well backfire, as a new controlled experiment suggests. Getting a second opinion can make software worse when the wrong agent is given the authority to rewrite the first one’s work.
Researchers tested Claude Opus 4.7 and Codex GPT-5.5 across 116 medium and hard Python tasks drawn from the LiveCodeBench benchmark. Codex passed 71.6% of the tasks alone, but reached 89.7% when Claude reviewed its answers – suggesting two heads are better than one there.
Flip the script, and things go haywire. While Claude scored 91.4% by itself, passing its output to Codex dragged the final pass rate down to 82.8%.
More like this
Different models, different personalities
The difference in performance was down to what happened when the code reviewers intervened. Claude repaired 26 failed Codex answers while breaking five that had previously worked – meaning it improved things in 21 instances. Codex fixed only three of Claude’s failures, while turning 13 correct solutions into incorrect ones.
“If you use Codex, which has worse accuracy than Claude, you actually make the overall accuracy worse than it was before,” says Zuodong Xiang, a co-author of the paper. He compares the pairing to human code review: asking a principal engineer to check a junior colleague’s work is useful, but putting the junior in charge of rewriting the principal’s code may not be.
Xiang says that engineering leaders therefore need to think about how their agents operate as a team, much in the same way they would with human employees.
“The big problem is that software is very dynamic. There is no unified solution now,” says Suzhen Zhong, a PhD student at Queen’s University in Canada who researches human-AI collaboration and agentic code review. She explains that, because different projects have different intentions, review cultures, and levels of complexity, individual benchmarks can struggle to properly reproduce what would happen in real life.
New York • September 15 & 16, 2026
The pace of change keeps accelerating. See what other leaders are doing about it, at LDX3 New York.
Strengths and weaknesses
Models may also have different strengths depending on the task. One agent that does well when tackling an isolated algorithmic problem could struggle to get its head around an organization’s architectural choices or coding conventions. Zhong suggests relying on public leaderboards alone is a bad way to tell developers which model they should deploy, and when.
That’s doubly so because of the costs involved. According to the paper’s own estimates, giving Codex output to Claude to double-check increased the average cost per task from $0.19 to $0.44 and latency from 38.5 to 112.4 seconds.
“If you have another agent reviewing the work, it also means there will be more usage and costs involved,” Xiang says. The price of improving code with a belt-and-braces check needs to be factored in by engineering leaders: smaller companies may find that the increase in accuracy is worthwhile for critical code, but harder to justify for everyday changes.
For companies that can afford it, Xiang advises to try out different kinds of agents to see how they work for you. “Using more than one agent at a time can sometimes be beneficial to your development process.”
It all means that, for engineering leaders, agent selection is far more than a procurement decision. The models may be artificial, but the hierarchy between them matters.
“You always want a more capable model to review your work,” Xiang says. “Just like with humans, you want someone more senior reviewing the work.”
