Did Claude just become too dumb to handle AgentOS complexity?

[mlamina]

This and other Reddit posts keep popping up about Claude's decrease in output quality and ability to follow instructions closely. When using it with Agent OS, it's almost become unusable for me. Even when my specs are perfectly clear, the implementation is mostly garbage. I feel like the only flow that still works somewhat reliably is "Plan Mode" -> "Build a simple thing".

I'm fully on board with AgentOS's engineering approach, but feel like Claude can't handle it anymore. Thoughts or suggestions?

[Bodekaer]

Same here. I need to repeatedly remind it of the commands.
Appears to happen consistently when context window grows too large especially, e.g. during spec creation, I'll reference relevant codebase, which seems to take away importance/attention to the detailed @Create-spec instructions.

[CasJam]

I've been noticing more unpredictability with Claude Code lately too. This is a big driving force behind Agent OS 2.0, which I'm currently working on...

As of right now, key improvements (among others) in 2.0 are:

• For Claude Code users, it's fully architected around subagents, which makes process adherence a lot more reliable, not to mention smarter/quality results.
• Better support for non-Claude Code coding agents (generic adaptor for any/all). These lose the benefit of subagents, but Agent OS 2.0 will work around that.

Coming soon!