# same task, same model: 40 seconds one run, 261 the next. my own "is it done?" check was rejecting the right answer — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/7dfe5898184b7098

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-11-01 18:27 ingested 2026-11-02 02:37

[original source ↗](https://www.reddit.com/r/AI_Agents/comments/1xlah7g/same_task_same_model_40_seconds_one_run_261_the/)

## retained text



```
i run a browser agent on a local 27B model (3-bit quant) on a single 3090. one of my test tasks is boring on purpose: search a korean bookstore for "python" and put the title, author and price of the top 3 books in a table. even after fixing three other things, three runs in a row took 40, 176 and 261 seconds. same prompt, same machine.

the first thing i blamed was thinking mode. the harness turns thinking on and off between steps, and every flip changes the start of the prompt, so the KV cache gets rebuilt. i fixed one case where the final answer flipped it and the model reread 77k tokens, then tried keeping thinking steady for whole runs. across 27 runs with different settings, the bookstore task still swung between under a minute and six minutes.

next i found our history cap was still sized for a 131k window while we run 192k, so it folded the transcript at about 40% and the model reread a lot. real bug, fixed, still slow. then the page extractor. the site has a hidden category menu with around 18,000 characters, and it beat the actual book list as the "main content". fixing that took one bad run from 373s to 124s, but that's one run, and the spread was still there.

what it actually was: before the model starts, a small rule layer turns the request into requirements the harness can check. it read "top 3 books, title, author, price" as "open 3 items and read each one". in the slow runs the model already had the full answer from the search results by step 5. my checker said not done, so it went and opened three book pages it never needed. in the logs that looks exactly like a confused model.

the fix was small. if every field asked for is something you can see in a list (author, price, star count and so on), the requirement becomes "collect from the list". before shipping i ran all 2,810 requests in our logs through the old and new rules. 14 changed, all the way i wanted, none the other way. after that the same task took 60, 64 and 44 seconds. 3 runs on one task, so i'm not claiming much, but the 4 minute runs stopped.

what i still haven't figured out is when to let the model think harder. the noise between runs was bigger than any difference between thinking settings, so i can't tell yet whether any of them help.

if you've chased run-to-run variance in your own agent, where did it end up coming from? the model, sampling, or your own scaffolding?
```



document fetched 2026-11-02 00:18 · canonical https://reddit.com/r/AI_Agents/comments/1xlah7g/same_task_same_model_40_seconds_one_run_261_the
