ARTFEED — Contemporary Art Intelligence

CAX-Agent: Lightweight Agent Harness for Reliable APDL Automation

publication · 2026-05-18

A recent publication on arXiv (2605.15218) presents CAX-Agent, a streamlined agent framework aimed at enhancing the dependability of MAPDL finite-element simulations through the use of large language models. This system tackles frequent failures stemming from unstructured execution by integrating domain-specific orchestration middleware that oversees tool lifecycles, workflow states, and fault recovery processes. CAX-Agent structures execution into three tiers: LLM service, agent harness, and solver backend. A recovery ladder progresses from deterministic rule adjustments to model-driven regeneration, context enrichment, and human intervention. The research assesses three recovery approaches—no_recovery, rule_only, and model_only—across 50 standard structural challenges.

Key facts

  • CAX-Agent is a lightweight agent harness for MAPDL automation.
  • It addresses reliability challenges in LLM-based finite-element simulations.
  • The system uses a three-layer architecture: LLM service, agent harness, solver backend.
  • A recovery ladder includes rule patching, model-driven regeneration, context enrichment, and human intervention.
  • Three recovery strategies were evaluated: no_recovery, rule_only, and model_only.
  • Evaluation was performed on 50 standard structural problems.
  • The paper is available on arXiv with ID 2605.15218.
  • The approach inserts domain-specific orchestration middleware.

Entities

Institutions

  • arXiv

Sources