Reconfigurable action theories enable dynamic agent behavior changes

Reconfigurable Action Theories: A Formal Foundation for Dynamically Composable Agent Harnesses

Artificial Intelligence

Summary

Most systems that describe how actions work assume the rules and tools stay the same. This paper presents a way to let actions change the rules themselves, so agents can switch tools and policies while running. The authors show how to keep behavior consistent even when parts of the system are added or removed. They also explain how to efficiently reason about these changing systems and compare their method to existing agent platforms.

What this means in practice

  • For robotics software engineers: Build adaptive robot control systems that safely switch tools and policies during operation by modeling changes within action theories.
  • For automation platform developers: Verify how adding or removing components impacts available actions and goals in complex agent frameworks using a formal dynamic action model.

Authors

Jianmin Ji

Abstract

Action theories typically fix their vocabulary and causal laws, whereas agent harnesses can change their tools, providers, and policies during execution. We introduce Reconfigurable Action Theories (RAT), in which actions can change the action theory itself. A theory module in RAT contains a vocabulary, causal laws, and an initializer. A RAT configuration records the active modules and the world state. Module loading and unloading are represented by actions governed by ordinary causal laws. RAT retains the standard causal action semantics of its language fragment when the active module set is fixed. We give sufficient conditions for at most one next configuration, preservation of existing behavior when a module is added, and equal results when modules are loaded in either order. We establish the computational complexity of reasoning tasks in RAT and translate finite RAT systems into nonmonotonic causal theories and Answer Set Programming (ASP), preserving bounded executions. Experiments compare finite RAT models with executions of official DeepSeek Harness, Pi, and OpenHands implementations. RAT provides a formal basis for checking how component changes affect a harness's available actions, behavior, and reachable goals.