My coding agent kept skipping confirmation when it decided the next step was obvious. Fixed it with hard gates.

I've been running a multi-agent setup on a real production codebase for a few months now, orchestrator plus an LLM doing most of the coding. The recurring issue: the agent would skip my confirmation step whenever it decided the next move didn't need approval. It wasn't hallucinating or broken. It just decided speed mattered more than waiting on me. A few times it had already edited three files before I even noticed. First fix I tried was the obvious one: tighten the prompt instructions, tell it explicitly to always stop and wait. Worked for maybe a day. Once the context window filled up enough, that instruction just stopped carrying weight. What actually worked was replacing soft instructions with structural gates. The agent has to produce something concrete, a written plan, an approval block, before it's allowed into the next phase. Spec, then plan, then execution, with a hard stop in between each one. No output, no progress. It's not optional anymore. Side effect I didn't expect: this also killed a second problem. Some of these agent runs were going 80+ hours and looked productive on the surface but were really just grep/diff loops going nowhere. The gates forced checkpoints where I'd actually catch that instead of letting it burn hours. Anyone else running agents in production hit this same wall? Curious if structural blocking is basically the only thing that holds up long-term, or if people found something else that works.