The most concerning thing here is the temperature problem. If your harness isn't providing deterministic output at a temperature setting of 0.0, it is broken.
Honestly, if you compare Tim's pre-GPT writing style, to his current stylometrics, you'll see a remarkably delta. It's as if he suddenly decided to change his established writing style and voice right when LLMs become widely available.
If you don't change your approach but just use CLI "intead of" MCP, you'll end up with a new spin on the same problems. The guardrails MCP provides (identity, entitlement, multi-principal trust boundaries) still need to exist somewhere.