The better LLMs get, the less you should over-scaffold them with agents — and the more you need to verify them. They're now good enough to pull a confident, happy-path "✅ it works" straight out of thin air (made to fit, rarely true), so lightweight security + diff-review matter more than ever — not less.
My quiet MVP isn't a clever agent, it's /stash: snapshot all progress, reset the session to refresh the KV cache and kill context rot. Then /remember runs a friction scan over your session logs — mining the moments you actually corrected the agent — and distills them into facts, episodes, and behavioral antigens that keep your CLAUDE.md fresh, so every new session starts already knowing your patterns. Build on primitives that help, and let focused subagents auto-discover instead of suffocating the model.
https://github.com/hamr0/li...