How do you keep AI coding agents from going off the rails on long tasks?

I'm nrslib, maker of TAKT (launching here Aug 9).

I built TAKT because agents kept forgetting my instructions in long-running work — implementing and reviewing blurred together, and I kept repeating the same feedback over and over.

My answer was to move the process outside the agent: plan → implement → review → fix loops defined in YAML, where reviews can't be silently skipped.

Curious how others handle this. Do you rely on prompts and rules files, a second agent as reviewer, or something else entirely? What actually worked for you?