AI coding agents are becoming useful, but they still burn budgets, loop on bad strategies, and finish without enough evidence. The next layer is trace intelligence, model routing, and control." AI Coding Agents Are Burning Budgets. The Next Layer Is Control. AI coding agents are getting better. They can read a repo, edit files, run tests, inspect errors, and try again. That is useful. But the problem showing up in real workflows is not just whether agents can write code. The problem is that agents can spend budget without producing finished work. They loop. They retry weak strategies. They switch files without explaining why. They chase unrelated errors. They claim completion without enough proof. And when the run ends, the human still has to ask: What actually happened? That is the gap the next generation of agent infrastructure has to solve. Not more autonomy first. Control first. The Problem Is Not Just Bad Code A bad patch is easy to see. A bad agent run is harder.…