Agentic Lybic: Multi-Agent Execution System with Tiered Reasoning and Orchestration

Abstract: Autonomous agents for desktop automation struggle with complex multi-step tasks due to poor coordination and inadequate quality control. We introduce Agentic Lybic, a novel multi-agent system where the entire architecture operates as a finite-state machine (FSM). This core innovation enables dynamic orchestration. Our system comprises four components: a Controller, a Manager, three Workers (Technician for code-based operations, Operator for GUI interactions, and Analyst for decision support), and an Evaluator. The critical mechanism is the FSM-based routing between these components, which provides flexibility and generalization by dynamically selecting the optimal execution strategy for each subtask. This principled orchestration, combined with robust quality gating, enables adaptive replanning and error recovery. Evaluated officially on the OSWorld benchmark, Agentic Lybic achieves a state-of-the-art 57.07% success rate in 50 steps, substantially outperforming existing methods. Results demonstrate that principled multi-agent orchestration with continuous quality control provides superior reliability for generalized desktop automation in complex computing environments.

• The paper introduces Agentic Lybic and its tiered FSM architecture for robust and adaptive desktop automation.
• It leverages a four-tier system comprising Controller, Manager, Worker, and Evaluator to manage multi-step task execution.
• Experimental results on the OSWorld benchmark show a 57.07% success rate in handling complex, multi-modal workflows.

Agentic Lybic: Multi-Agent Execution System with Tiered Reasoning and Orchestration

The paper introduces Agentic Lybic, an FSM-based multi-agent execution architecture for desktop automation. By leveraging a sophisticated orchestration framework, the system achieves efficient task execution even in complex multi-step scenarios while maintaining high reliability through continuous quality control.

System Architecture

Agentic Lybic's architecture is built on a four-tiered framework comprising a Controller, a Manager, a Worker subsystem, and an Evaluator. This design adopts finite-state machines (FSMs) to manage dynamic transitions, allowing for adaptive strategies contingent on task complexity and execution outcomes.

Figure 1: State transition diagram of Agentic Lybic showing the tiered orchestration workflow.

Central Controller

The Controller operates as the core orchestrator, managing state transitions and ensuring coherent interaction between components. It defines situations like REPLAN, SUPPLEMENT, GET_ACTION, QUALITY_CHECK, FINAL_CHECK, and EXECUTE_ACTION, each progressively refining task execution strategies. State transitions are determined by a comprehensive state space model $S = (S_T, S_{ST}, S_E, C)$, enabling granular process control.

Manager and Worker Subsystem

The Manager handles task decomposition using a DAG-based representation, facilitating parallel task execution and adaptive re-planning through triggers. The Worker subsystem includes Operators for GUI tasks, Technicians for system operations, and Analysts for decision support, each utilizing role-specific capabilities to execute actions effectively.

Evaluator

The Evaluator provides continuous quality assessment and early error detection, implementing a Gate Decision Framework with triggers for periodical checks, stagnation, and task completion. This proactive approach to error handling through adaptive feedback enables robust and efficient execution.

Experimental Evaluation

The system's capabilities were evaluated on the OSWorld benchmark, a comprehensive desktop automation testbed consisting of 361 tasks spanning various applications.

Figure 2: OSWorld benchmark task complexity analysis.

Agentic Lybic achieved a new state-of-the-art performance with a 57.07% success rate in 50 steps, surpassing prior systems like CoAct-1 and Agent S2.5. The results underscore its robustness in handling complex workflows, particularly in applications requiring multi-modal coordination.

Figure 3: Successful case demonstrating multi-modal task execution.

Robustness and Limitations

Agentic Lybic's robustness is attributed to its adaptive orchestration mechanisms that manage diverse execution scenarios efficiently. However, certain benchmark evaluation limitations, such as rigid criteria leading to misclassification of successful executions as failures, were noted (Figure 3). The system's execution model accommodates shifts in task demands, ensuring effective error recovery and consistent task fulfillment.

Figure 4: Failure case analysis demonstrating evaluation standard limitations.

Conclusion

Agentic Lybic significantly contributes to advancing desktop automation by integrating an FSM-based multi-agent system that enables real-time adaptation and robust error management. Its architecture not only boosts task success rates but also highlights the potential of organized multi-agent orchestration in solving diverse automation challenges in computing environments.

Future research may explore enhanced modalities, particularly in domains demanding real-time updates such as video manipulation and dynamic user interfaces, expanding the applicability of this architecture to scenarios necessitating collaborative interactions or distributed system coordination.