EnCompass: Enhancing Agent Programming with Search Over Program Execution Paths

2025-12-03

AI Summary

Existing LLM agent programming approaches tightly couple workflow logic with runtime reasoning strategies (e.g., tree search), impeding independent modeling and substitution of either component. Method: We propose a decoupled agent programming paradigm centered on the “Probabilistic Angelic Nondeterminism” (PAN) model, which formally separates deterministic workflow logic from pluggable, strategy-agnostic reasoning policies. Based on PAN, we design EnCompass—a Python framework that automatically compiles annotated agent programs into searchable execution-path spaces via decorator-based instrumentation. Contribution/Results: Evaluated across three diverse tasks, EnCompass enables flexible, low-effort strategy switching (e.g., DFS, BFS, MCTS) with minimal code changes—yielding substantial improvements in task success rate and robustness. This work establishes a foundation for modular LLM agent design, systematic strategy experimentation, and reliability-oriented optimization.

Technology Category

• Cognitive Modeling & Cognitive Systems: Agent Architectures
• Multiagent Systems: Modeling other Agents
• Reasoning under Uncertainty: Probabilistic Programming

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Agentic search
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

We introduce a new approach to agent programming, the development of LLM-based agents. Current approaches to agent programming often entangle two aspects of agent design: the core workflow logic and the inference-time strategy (e.g., tree search). We introduce"probabilistic angelic nondeterminism"("PAN"), a programming model that disentangles these two concerns, allowing the programmer to describe the agent workflow and independently experiment with different inference-time strategies by simply changing a few inputs. We provide an implementation of PAN in Python as the EnCompass framework, which uses a Python decorator to compile agent workflow programs into a search space. We present three case studies that demonstrate how the framework lets the programmer quickly improve the reliability of an agent and easily switch between different inference-time strategies, all with little additional coding.

Problem

• Separates agent workflow logic from inference strategies
• Enables flexible experimentation with different search methods
• Improves agent reliability with minimal additional coding

Innovation

• PAN disentangles workflow logic from inference strategies
• EnCompass compiles agent programs into search spaces
• Framework enables quick reliability improvements with minimal coding