Intel / video

Benchmarking Coding Agents on New vs Legacy Codebases — Denys Linkov, Wisedocs

Source: youtube.com
Author: AI Engineer
Date: Published Aug 8, 2026

Context

Generating a plausible replacement for a large codebase can conceal missing functionality. In Wisedocs' refactoring case study, a medical-document processing team asked GPT-5.5 at its extra-high setting to carry out a repository consolidation as a single goal. The reported completion in about ten minutes with roughly 2,000 lines of code, but inspection showed it had produced scaffolding while leaving out the underlying models.

The team's own refactor, which came earlier, ran differently: the speaker says every pull request received human review, with some local -based checks, and that 15 of 17 requirements were met. The speaker also reports faster processing, lower cost and quicker feature delivery after the rebuild, but those are self-reported. The experiment establishes a narrower point: an agent's completion message needs to be checked against the functionality it was asked to preserve.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agent skill — A reusable instruction file that teaches an agent how to do one job well — the procedure, the tools, and what counts as done.

Key quotes

“it completed its goal in in 10 minutes and 22 seconds. And it only wrote 2,000 lines of code, which was a little bit fishy.”
Denys Linkov

“it's very easy to undergo AI psychosis, where you look at a deep research report that's 20 pages long”
Denys Linkov

“taking on technical debt and refactoring later is getting exponentially easier as the days go by”
Denys Linkov

“we can ship features that would take multiple months in under a week”
Denys Linkov

“sometimes it's good to pause, build a monorepo, and forge ahead”
Denys Linkov

Read the source www.youtube.com