← sources

tested three coding agents on the same real codebase. the biggest surprise was how much of your setup follows you between them

source: reddit (reddit:AI_Agents)published 2026-10-02 16:18ingested 2026-10-03 02:35

retained text

gave the same feature spec to cursor, codex and claude code on my own live saas. same branch, cheapest paid plan each, default settings, 16 point acceptance list written before the first run

thought i was testing the agents. half of what i found was about the layer around them

cursor imported my claude code plugins without asking. there's a toggle for it, on by default. only caught it because the agent reached for a plugin i never installed there. so my first run was measuring my old setup in a new window

then i bought a brand new claude account specifically to get a clean run. didn't help. plugins, skills and an mcp server pointing at a completely unrelated repo all loaded anyway, because they live in the home folder not the account. only clean session i got was CLAUDE\_CONFIG\_DIR pointed at an empty dir. worth knowing if you ever want to measure anything, or if you're wondering where your context goes before you type a word

the other thing that got two of three: your project remembers stuff that isn't in your instructions file. in my case the list of generated docs is hardcoded in four places and one of them says don't reference anything outside it. two agents added a required doc and never wired it in, so the file sits in the repo and nothing ever opens it. generation reported success. the done-when check passed. i only caught it by opening the generated CLAUDE.md by hand

what actually separated them was scope of responsibility, not model quality. my local db was down during the runs. cursor wrote "couldn't verify anything" and stopped. codex asked to start docker itself, applied migrations, wrote playwright tests and found a serialisation bug in its own code. claude code did the same plus wrote a contrast test across design presets it had just invented and fixed two that failed wcag aa

numbers: claude code 38 min and 6% of a weekly limit, cursor 52 min and 3% of a monthly quota, codex hit its 5h cap mid feature and needed a 3.5h wait

caveat, one run each. someone pointed out i should've run one tool twice to get the noise floor first, which is fair and i'm doing that next time

writeup with the spec, screenshots and three live demo apps in the comments

document fetched 2026-10-03 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1wvy1lo/tested_three_coding_agents_on_the_same_real