/AI15h ago

David Manheim, ALTER founder, finds Codex 5.5 bottlenecks multi-agent codebase refactoring by sequentially concentrating control in a single agent

Boaz Barak notes a config parameter limits active subagents.

--0--
Comments
Original post
Boaz Barak@boazbaraktcs#133inAI

@davidmanheim @jxnlco do you have suggestions?

David Manheim@davidmanheim

@boazbaraktcs So 5.5 is better at this than last time I tried, where it didn't actually delegate. But when I cleared session and asked it to continue, with same "prefer low cost agents"; it launched 5.5 agents instead of 5.4-mini subagents it used before; now I'm out of 5-hour window credit.

7:15 AM · Jun 2, 2026 · 3 Views
Sentiment
Sentiment building, check back later.
Cluster Engagement
-
Views
-
Comments
-
Reposts
-
Bookmarks
Expand data
Posts from X
Most Activity
Most ActivityTimeline
VIEWS26REPLIES1
Boaz Barak@boazbaraktcs

@davidmanheim There is a max thread parameter in config.toml with how many agents it is allowed to launch. TBH I think codex works much better from the app compared to the CLI, and there you can track all the subagents.

David Manheim@davidmanheim

@boazbaraktcs To be clear, I'm having it refactor a too-large main.ts into different features, so maybe it's just being smart keeping single agents in control, but I need to micromanage too much here.

And weird; it doesn't know how many agents it started already / it is allowed to launch?

12hViews 26Likes 0Bookmarks 0