An orchestration graph cycles between two agents without task progress. Which safeguards should be added?

1. Step and token budgets, termination criteria, and trace-based failure analysis
2. Permanent retries with no progress signal
3. A larger unbounded context window as the only control
4. Broader tool privileges