Decoupled Planning and Execution for Trustworthy Agents in Closed-Environment

Enterprise systems rely heavily on LLM agents with typically large models(with parameters) to execute multi-step procedures. They use a dynamic approach; first the agent creates a plan and uses it to execute the input query and generate output. These agents do not require any offline preparation, yet they still fail in producing consistent results despite using a powerful model, rendering them unreliable. We presume that dynamic re-planning not only causes structural drifts, but also amounts to repeated redundant token costs. Existing literature points towards Localized task decomposition bounds updates within sub-tasks, complementing prior efforts that employ graph constraints, delta patching, and topological separation to safeguard global dependencies to mitigate the shortcomings, which motivates us to study a hybrid approach of query resolution that involves decoupling of planning and execution paths and employing DAGs(Directed Acyclic Graphs) to represent pre-defined workflows. We conducted two studies in closed environment setup. First, we study two different systems across a custom benchmark and do a cost analysis; one uses traditional dynamic approach and other one uses the hybrid approach. Then we study the novel hybrid system against three graph compiler models of different sizes. Results obtained from these studies suggest that the hybrid approach is more reliable, accurate and cheaper and that smaller compiler model performs fairly in ingestion phase, hence supporting central claim that small model executing a well-structured, pre-compiled plan is cheaper and more reliable than dynamic re-planning.