Unfortunately, I’m stuck with Claude Code, which is my least-favourite harness. It’s tree navigation (or lack thereof) is such a dumpster fire. 
Also, I’m currently working in a legacy codebase that lacks good guardrails for Claude to bump off of. This results in a lot of ad-hoc verification, and inevitable drift as the agent leverages certain half-truths and assumptions into outright fabrications (this compounds between sessions). The durable facts that often ground the model just aren’t present in this codebase at the moment. (I’m working on it…)
The old quote rings true: “Information wants to be free.”
I do love the power with which I can turn ideas into working prototypes. But man, I just can’t wait until the dust has settled, and LLMs become a “boring” technology.
