Critique loop
Updated 2026-07-11
Questions this answers
- How do I get higher-quality output without just hoping the first try is good?
- Is it better to run many agents at once or to review more carefully?
- My AI's first draft looks fine but has subtle mistakes, how do I catch them?
- Where should I spend my best, most expensive model?
The Fix
Here's a move that consistently improves output: run a critique loop. One model does the work, then a different model reviews and critiques it before you accept it. The reviewer catches things the author missed, the same way a second set of human eyes does.
Shopify's CTO shared that this beat the alternative people reach for first, which is running a pile of agents in parallel and hoping one nails it (the slang for that is tokenmaxxing). The parallel approach actually produced worse results. The critique loop was slower, and yes people complain about the wait, but the output was better.
There's a companion habit worth naming: watch your generation-versus-review ratio, and spend your biggest, slowest model on the review, not the first draft. Reviewing well is where the quality actually comes from.
When to Use It
Use a critique loop whenever the output matters more than the speed: important writing, code you're going to ship, analysis someone will act on. Skip it for quick, low-stakes, throwaway work where a first draft is fine and the extra latency isn't worth it.
In the Wild
Been hearing wild stuff from folks inside big companies lately. Promotions, firings, and perf reviews are getting decided by tokens consumed and skills/MCPs connected. That’s the metric. That’s how they’re deciding who’s “good at AI.” It gets worse. People are literally running Show more
On getting better results by having models review and iterate rather than accepting a first pass.