Runs happen inside the harness and write to a log nobody opens. You find out an agent got it wrong from a customer, not from a test.

An agent crashes, times out, or loses its machine. Nothing announces it. The ticket still says an agent is working on it, three days later.

An agent needs a decision only a person can make, so it posts in Slack. Whoever is awake replies. Nobody owns it, nobody logs it, and the next agent never learns what was decided.

You cap how many run at once, so agents working on unrelated files sit idle. The two that are editing the same file still run straight into each other.

One task has to finish before another can start, but nothing records that. So agents build against a half-finished migration, and someone orders the board by hand every morning.

One flat store means each agent loads all of it, every time. It costs more, it gets slower, and the more you add the less the agent finds what matters.