CC as orchestrator is structurally unreliable — harness instructions are reset every session

Summary

Claude Code (CC) is marketed as an orchestrator for multi-agent harness workflows. In practice, CC resets its behavioral patterns every session and defaults to doing code work directly (Edit/Write tools) instead of delegating to sub-agents — even when explicitly and repeatedly instructed not to.

What happened

• User built a multi-agent harness over 3 months with CC as orchestrator
• Explicit rule established: CC must delegate code file changes to agents, never use Edit/Write directly on .py/.html/.js/.ts etc.
• Every new session: CC violates this rule and starts editing code directly
• User has to correct CC repeatedly within every session
• Rule was written into CLAUDE.md, added as hooks, added to memory files — none of it reliably prevents the behavior

Root cause

CC has no persistent behavioral memory across sessions. Instructions in CLAUDE.md are "suggestions" CC can rationalize away. The model's default behavior (do the task directly) overrides user-defined orchestrator constraints every time.

Impact

• 3 months of harness architecture work undermined by CC's inconsistent behavior
• User spent hours per session re-correcting the same behavior
• Work attributed to wrong causes (blaming sub-agents) when the real issue was CC's behavior

What should change

1. Anthropic should clearly document that CC cannot reliably maintain role constraints across sessions
2. If CC is sold as an orchestrator, session-persistent role enforcement must be a feature, not a user responsibility
3. Consider a hard mode where CC literally cannot use Edit/Write on specified file types

Severity

High — this affects any user trying to build serious multi-agent workflows with CC as orchestrator.