What's this about?
Two days after launch, Opus 5 hit an "Elevated Errors" incident, and the status-page thread filled with paying users venting about reliability — dropped sessions, model-switching bugs, HTTP errors. Separately, Redditors digging through Claude Code found a hardcoded instruction telling Opus 5 specifically not to use subagents.
What exactly is a "subagent"?
A subagent is a helper AI that the main AI spins up to work on a piece of the task in parallel — like a contractor hiring subcontractors. A hardcoded "don't hire subcontractors" note aimed at one specific model usually means the vendor found that model misusing them (too eager, too expensive, or error-prone) and patched the behavior in the prompt rather than the model.
What's the community saying?
92 + 26 points, 76 + 13 comments
- "Fix the infra before shipping models" — Team subscribers threatening to churn; one quips that a company selling an AI developer still needs humans to fix its own systems.
- On the subagent find: is this a temporary launch guardrail or a sign Opus 5's agent-coordination is undercooked?
- Amusement at prompt archaeology becoming a genre: the same week Claude Code brags about deleting 80% of its prompt, users reverse-engineer what remains.