Why does Claude Code seem to get worse over a session, and should it auto-spawn sub-agents?

A frequent, fixable cause is uncontrolled sub-agent spawning. When an agentic coder auto-delegates every task to sub-agents, three things go wrong: sub-agents report "Done" without validation, they lose the main thread's context, and you get rework loops where the main agent re-does sloppy work. The felt experience is the tool "getting dumber" over a session, even on a capable model.

The fix that worked (Opus-class model, real project): one global rule in the project's instructions file - no sub-agents without express approval. That keeps the main agent in control: it does the work directly, or, for genuinely parallel bulk tasks, sets up the process and validation first, then delegates deliberately. Performance returned immediately. Sub-agents are still useful - the point is that the main agent should decide when to use them, not spawn them reflexively.

Full write-up: https://www.tigzig.com/post/claude-code-sub-agents-performance-fix. Related: the rules for working with AI coders https://www.tigzig.com/agents-faq/what-rules-to-follow-working-with-ai-coders and the Anthropic 400k-session study on what actually helps https://www.tigzig.com/agents-faq/anthropic-400k-claude-code-sessions-study.