ARTFEED — Contemporary Art Intelligence

COVENANT: Compiler for Aligned LLM Agent Execution

ai-technology · 2026-07-29

Researchers propose COVENANT, a compiler-and-interpreter architecture to ensure large language model (LLM) agents follow natural-language workflow instructions faithfully. The system treats instructions as source programs, converting them into a workflow abstract syntax tree (WAST) and then a workflow control-flow graph (WCFG). A runtime controller enforces step-by-step compliance, preventing agents from skipping steps, taking unsupported branches, or executing valid steps with unsupported arguments. The work addresses a failure mode called workflow misalignment, where agents deviate from specified procedures when instructions are supplied as prompt context. The paper is available on arXiv under identifier 2607.25400.

Key facts

  • COVENANT is a compiler-and-interpreter architecture for LLM agents.
  • It converts workflow instructions into a workflow abstract syntax tree (WAST).
  • The WAST is lowered to a workflow control-flow graph (WCFG).
  • A runtime controller ensures aligned execution.
  • Workflow misalignment occurs when agents skip steps or take unsupported branches.
  • The approach treats instructions as source programs, not prompts.
  • The paper is on arXiv with ID 2607.25400.
  • The work focuses on natural-language workflow instructions like retail-payment policies.

Entities

Institutions

  • arXiv

Sources