COVENANT: Natural-Language Workflow Compilation for Aligned Agent Execution

2026-07-28

AI Summary

This work addresses the prevalent issue of procedural misalignment—such as step skipping, privilege escalation, or parameter violations—in large language model (LLM) agents executing natural language workflows. To mitigate this, the study introduces, for the first time, concepts from programming language compilation into this domain, treating natural language instructions as source code. It proposes a compiler-interpreter architecture that constructs a Workflow Abstract Syntax Tree (WAST) and a Workflow Control Flow Graph (WCFG) from the input instructions. A runtime controller then executes the workflow node-by-node while performing dynamic validation to ensure strict adherence to the original specification. Experimental results on 120 test cases demonstrate that this approach increases task success rate from 50.00% to 83.33% and reduces procedural misalignment incidents by 62.75%.

Abstract

Large language model (LLM) agents are increasingly entrusted with natural-language workflow instructions (e.g., retail-payment policies) that specify not only what outcome to achieve, but also which steps, branches, and tool interactions are permitted. When these instructions are supplied as prompt context, however, the model retains control over both procedure selection and step execution. As interactions accumulate, an agent can skip required steps, take unsupported branches, or execute a valid step with unsupported arguments or effects--a failure mode we call workflow misalignment. In this work, we propose COVENANT, a compiler-and-interpreter architecture for workflow-aligned agent execution. Our key insight is to treat workflow instructions as source programs rather than prompts. COVENANT converts the instructions into a workflow abstract syntax tree (WAST) and lowers it to a workflow control-flow graph (WCFG). At runtime, a controller interprets the WCFG one node at a time, checks each proposal against requirements extracted from the instructions before committing controller state or advancing the graph, and returns diagnostic feedback for repair. To evaluate COVENANT, we use 120 cases from three existing benchmarks, spanning seven workflow scenarios. Compared with state-of-the-art LLM agents, COVENANT improves benchmark success from 50.00% to 83.33% and reduces the workflow-misalignment failure rate from 42.50% to 15.83% (62.75% relative). These results show that COVENANT substantially mitigates workflow misalignment, moving LLM-agent alignment beyond isolated prompt following toward reliable execution of complex and multi-step workflows.

Problem

• workflow misalignment
• natural-language workflow
• LLM agents
• instruction following
• workflow compliance

Innovation

• workflow alignment
• natural-language compilation
• control-flow graph
• LLM agent
• misalignment mitigation