Developers presented the HEXIS framework for controlling the sequence of actions of agents

The development team has released HEXIS — a tool designed to manage the behavior of AI agents through formalization of their skills. The system converts a set of agent functions into extended finite state machines (EFSM), which allows for strict ordering of operation execution and eliminates logical errors in action chains.

The main task of HEXIS is to ensure the predictability of autonomous systems. The framework supports dynamic updating of models based on trace analysis, which allows for adjusting agent logic without a full system rebuild. The toolkit includes a set of ready-made examples of automata for various use cases.

To ensure reliability, a hermetic test suite is included in the repository. This allows developers to test the correctness of agent operation in an isolated environment, guaranteeing that changes to the logic do not disrupt established interaction protocols. The project is focused on creating stable workflows in complex multi-step tasks.

Using finite automata to manage AI agents allows for limiting the state space in which the model can exist. This reduces the likelihood of hallucinations or execution of invalid command sequences, which is critical for automating business processes and working with external APIs.