Orbit · Explore

AI Lab · by Youssef Mahmoud · 2026-10-16

Are you actually trusting AI coding agents with multi-file refactors yet?

I've been letting an agent handle refactors that touch 10+ files for a month. It's great when tests are solid and a disaster when they aren't. Where is your line?

47 sparks · 13 comments

Comments

Luca Bianchi · 14 likes

Be careful with secrets in context, people forget that part.

Sanne de Vries · 8 likes

Tests are the whole game. With good coverage I let it run wild, without it I review every line.

Léa Bernard · 10 likes

Same. Coverage is basically the permission slip.

Paper Orbit · 8 likes

Honestly I just git stash before every run. Anxiety-driven development.

Oliver Smith · 13 likes

lol the real best practice

Chloe Davis · 3 likes

Cost matters too. Long agent loops burn tokens fast and half the time it re-reads the same files.

Thandi Mokoena · 9 likes

Caching helps a lot, check your provider settings.

Youssef Mahmoud · 2 likes

I trust it for renames and API migrations, not for anything with business logic.

Silver River · 8 likes

Renames are the easy win, agreed. Logic is where it gets creative in a bad way.

Avery Johnson · 2 likes

Review time ate my savings. Diffs from agents are huge and noisy.

Chiara Romano · 2 likes

Ask it to commit in small steps, it helps.

Curious Runner · 2 likes

Enthusiast here: it already beats my junior-me on boring migrations.

Cedar Runner · 3 likes

Low bar but fair :)