Skip to content
Search
paperApril 2026Unreviewed

Harness as an Asset: Enforcing Determinism via the Convergent AI Agent Framework (CAAF)

Tianbao Zhang

View resourcePDF

Abstract

Large Language Models produce a controllability gap in safety-critical engineering: even low rates of undetected constraint violations render a system undeployable. Current orchestration paradigms suffer from sycophantic compliance, context attention decay, and stochastic oscillation during self-correction. We introduce the Convergent AI Agent Framework (CAAF), which transitions agentic workflows from open-loop generation to closed-loop fail-safe determinism via three pillars: (1) Recursive Atom

Categories

Agentic Threats

Cite

@misc{zhang2026harness,
  title = {{Harness as an Asset: Enforcing Determinism via the Convergent AI Agent Framework (CAAF)}},
  author = {Tianbao Zhang},
  year = {2026},
  month = apr,
  eprint = {2604.17025},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2604.17025}
}