HEXIS: Compiling Skills into Extended Finite State Machines

Problem

Agents currently face challenges in correctly applying skills and inferring subsequent operations, leading to inefficiencies in task execution. This paper addresses this gap by proposing a structured approach to skill management through the use of extended finite state machines (EFSM). The work is presented as a preprint and has not undergone peer review.

Method

The core technical contribution is the HEXIS framework, which compiles agent skills into EFSMs, effectively separating knowledge from control flow. Key components of the method include:

• Skill Compilation: Skills are transformed into EFSMs, allowing for a clear delineation between knowledge representation and operational control.
• Local Instructions: These guide reasoning and generation within specific states of the EFSM, enhancing the agent's ability to make informed decisions based on its current context.
• Execution Tracking: The framework records execution progress and intermediate results, providing a mechanism for monitoring the agent's performance.
• Transition Conditions: Explicit conditions are defined to determine the next operations based on the current state, facilitating more accurate decision-making.
• Incremental Compiler: This component maps skill clauses and tool interfaces to state operations, local instructions, data bindings, and transitions, ensuring a comprehensive integration of skills into the EFSM framework.
• Development Trace Alignment: The method aligns execution traces with existing states to identify any missing operations and dependencies, which is crucial for maintaining operational integrity.
• State Updates: The framework incorporates updates following static checks and the replay of traces, ensuring that the EFSM remains current and effective.

Results

The implementation of HEXIS demonstrates significant improvements in operational efficiency:

• Success Rate Improvement: A notable increase of 16.1 percentage points in success rates was observed when compared to the baseline method Skill + ReAct.
• Execution Token Reduction: The framework achieved a reduction in execution tokens ranging from 38.4% to 88.9% across various benchmarks, specifically with the Qwen3.8-27B model.

Limitations

The authors do not report any limitations in the study. However, the absence of peer review may imply that the findings should be interpreted with caution until validated by the community.

Why it matters

The HEXIS framework has significant implications for the development of more capable and efficient AI agents. By structuring skill management through EFSMs, it enhances the agents' ability to reason about their actions and adapt to changing conditions, paving the way for more sophisticated applications in autonomous systems and robotics.