Notes on building Super Terminal — what works, what doesn't, and what we're changing.
CodeRabbit, Greptile, Qodo and GitHub Copilot reviewed honestly. They all aim at the pull request, which is the wrong moment for code an agent wrote ten seconds ago against rules it only partly followed.
The tools for running several agents at once solve a real problem well — and it is a throughput problem, not a trust one. Which of the two you actually have decides what to install.
The three agents are closer than the marketing suggests. What separates them is the surface you work in and the file your rules live in — and on the thing people actually worry about, all three behave the same way.
Two identical buttons, one instruction, and both of them gone. A working account of why agents overreach, why your rules stop at the vendor boundary, and which guardrails actually held — plus an honest look at the alternatives I tried first.
AGENTS.md won the format war, and Claude Code still doesn't read it. But the filename was never the hard part: every one of these files is text added to a prompt, not configuration a tool enforces — and the best model in one benchmark followed fewer than 30% of instructions completely.