I just moved my work into the Projects so I could have access to chats across desktop/laptop work. However, since I have moved, Claude Opus 5 has been very hard to work with and the work it passes back and forth to other agents is simply bad and wrong most of the time. I’ve had to create a rule that it is required to check each agents work that it hands off to and about 50% of the time, it finds errors. I never had this poor of quality with Opus 5 in Workspaces. I was hopeful Projects would be something that would solve my problems, but it has only created them. In addition, my token usage has exploded. I’ve heard Claude has had some issues this week so I’m coming here to see if others are having issues with Projects or perhaps this is a Claude issue.
This is mostly how Projects works rather than Opus 5 getting worse. The main chat is a coordinator that hands each task to separate agents, and each agent only gets the brief for its task plus the Project’s shared context, not your full chat history (docs). That’s why handoffs can drop detail, and why usage climbs (several agents run in parallel, each billing its own work). We’re not seeing anything unusual with Opus 5 on our end, so I wouldn’t put this on the model.
What helps:
- Put the full context each task needs (files, constraints, what “done” means) into the request, instead of relying on the coordinator to re-check work afterward.
- Keep agents on your model: add a line to your Project rules like “Start every agent on the same model as this Project unless I ask otherwise.” It’s guidance, not a hard setting, but it usually holds, and firmer control is on the team’s radar.
- For a small, single-file change, a regular chat or Cloud Agent is faster and cheaper. Projects shine when you have several tasks running at once.