The hype cycle for AI agents often fixates on model intelligence, but Mikhail Liublin’s work on NoodleTomato exposes the true chokepoint: architectural resilience. Liublin is developing an agent system for AI video generation that operates over hours rather than seconds. The project takes a user brief and orchestrates a complex pipeline of planning, scriptwriting, and delegation. This multi-step autonomy shifts the engineering challenge from prompt engineering to system design.

Beyond the Six-Second Clip

NoodleTomato does not generate a single six-second clip. It runs for hours, managing a workflow that mimics a production team. Liublin’s approach demonstrates that long-running agents require robust state management to handle interruptions, retries, and context persistence. The model acts as a component within a larger distributed system, not the sole driver of success. When the agent delegates work, it must track dependencies across multiple sub-tasks, a problem that pure LLM capabilities cannot solve.

Orchestration Is the New Frontier

The core insight from Liublin’s development process is that the 'hard part' is not the model. It is the infrastructure that keeps the agent coherent over time. As agents move from chatbots to autonomous workers, the complexity of their state machines explodes. Developers must build reliable mechanisms for memory, error recovery, and task prioritization. Without these, even the most advanced model will fail when a task spans hundreds of steps or multiple API calls.

Key Takeaways

  • Long-running agents require stateful architecture, not just powerful models.
  • NoodleTomato demonstrates hours-long autonomy through delegation and planning.
  • The bottleneck for agent reliability is orchestration and error handling.

The Bottom Line

Stop blaming the model for your agent’s failure. If your system can’t maintain coherence over time, you don’t have an AI problem. You have an engineering problem. The industry needs to pivot from chasing higher benchmarks to building resilient agent frameworks. Liublin’s work on NoodleTomato is a blueprint for this shift. We need tools that manage complexity, not just generate text. The future of autonomous agents lies in their ability to persist, adapt, and recover. That is a systems engineering challenge, and it is the only one that matters.