Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2608.26197

Harness Engineering for Predictable Agentic Systems: An Empirical Study of Deterministic Execution Constraints

Saransh Dhage
View original source ↗Read source document ↗Export metadata ↓

Abstract

Large Language Model (LLM) based agents exhibit substantial run-to-run execution variance even when given identical tasks and tools -- acceptable for exploratory use but unacceptable in regulated domains such as finance and compliance. We study harness engineering: wrapping an agent in a deterministic execution layer (finite-state control, forced tool selection, output validation, bounded retry, and structured planning) and measuring its effect on execution determinism and task success. Across two synthetic tasks (finance and legal) and two open-weight models (Qwen-2.5-7B-Instruct, Gemma-3-27B), a first-pass harness produces a mixed result: it significantly improves reproducibility in one of four model-task cells, significantly degrades it in two, and has no effect in the fourth. A trace-level diagnostic finds the cause: once tool sequence, state sequence, and output are already highly consistent, an unconstrained free-text planning step becomes the dominant remaining source of variance. Adding Structured Planning -- validating the plan against a fixed schema before any tool is invoked -- eliminates the effect entirely: three of four cells reach a Reproducibility Rate and Determinism Index of 1.000 at N=100, and task success rises to 100% in three of four cells. The gain is not free: token cost falls in every cell, but latency shows a genuine, sample-size-robust split by model -- one becomes faster under the constraint, the other markedly slower. We argue harness engineering is a distinct, effective discipline for agent reliability, but its cost is model-dependent and must be measured, not assumed.

Explore related subjects

cs.SE

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Saransh Dhage. 2026-08-25. Harness Engineering for Predictable Agentic Systems: An Empirical Study of Deterministic Execution Constraints. https://arxiv.org/abs/2608.26197

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Search-based Trace Diagnostic for Cyber-Physical Systems

Cyber-physical systems (CPS) development requires verifying whether system behaviors violate their requirements. This analysis often considers system behaviors expressed by execution traces and requirements expressed by signal-based temporal properties. When an execution trace violates a requirement, engineers must solve the trace diagnostic problem---they need to understand the cause of the breach. Automated trace diagnostic techniques aim to support engineers in the trace diagnostic activity. This paper proposes search-based trace diagnostic (SBTD), a novel trace diagnostic technique for CPS requirements. Unlike existing techniques, SBTD relies on evolutionary search. SBTD starts from a set of candidate diagnoses, applies an evolutionary algorithm to generate new candidate diagnoses (via mutation, recombination, and selection), and uses a fitness function to determine the qualities of these solutions. Then, a diagnostic generator step is performed to explain the cause of the trace violation. We implemented Diagnosis, an SBTD tool for signal-based temporal logic requirements expressed using the Hybrid Logic of Signals (HLS). We evaluated Diagnosis by performing 34 experiments for 17 trace-requirement combinations for property violations. We assessed the effectiveness of SBTD in producing informative diagnoses and its efficiency. Diagnosis achieved expert-aligned diagnoses for 29/34 experiments and scaled to the full HLS benchmark, whereas state-of-the-art literature remained restricted to a subset due to performance and language limitations. SBTD treats trace-checking as a black box, which makes the checker replaceable. Substituting our HLS checker for an STL monitor, e.g., RTAMT, reproduces on two requirements the diagnoses at two to three orders of magnitude lower per-check cost.

cs.SE↗
arXiv2026

Toward Secure Code Generation: Bridging Correctness and Security via Task-Adaptive Vulnerability Modeling and Execution-Based Benchmarking

Large language models (LLMs) are increasingly used for program synthesis, yet they often generate code that is functionally plausible but insecure. Progress in secure code generation has been hindered by benchmarks that are small, non-executable, leak mitigation details, or rely on noisy analyzers and subjective judgments, making it difficult to measure whether security improves without sacrificing correctness. We address these gaps with CodeSecEval, an execution-based benchmark for secure code generation, comprising 255 Python tasks spanning 77 CWE categories. Each task provides paired insecure and secure implementations together with executable functional and vulnerability-targeted security tests, enabling precise and reproducible evaluation of secure code generation and insecure-code repair. Building on CodeSecEval, we propose SecAwareCoder, an agent-based framework that shifts code generation toward secure-by-construction synthesis. SecAwareCoder performs task-adaptive threat modeling to identify security-sensitive regions and derive task-grounded vulnerability hypotheses, uses these hypotheses to guide both constraint-aware code generation and security-aware test synthesis, and leverages execution feedback for targeted refinement. Experiments across multiple LLM backbones show that SecAwareCoder consistently improves Pass@1 and security robustness over prompting and analyzer-driven baselines, narrowing the security--correctness gap in LLM code generation.

cs.SE↗
arXiv2026

Enabling Communication via APIs for Mainframe Applications

Mainframe systems continue to support critical applications across industries such as banking, retail, and healthcare. Exposing their functionality through Application Programming Interfaces (APIs) enables reuse and development of new applications, but identifying and implementing APIs for legacy code remains challenging. It requires understanding complex programs, separating dependent components, introducing new artifacts, and preserving functionality and Service Level Agreements (SLAs) such as Turnaround Time (TAT). We propose a framework for APIfication of legacy mainframe applications. Candidate APIs are identified from artifacts such as transactions, screens, control-flow blocks, inter-microservice calls, business rules, and data accesses. Static analyses, including liveness and reaching definitions, are then used to traverse the code and automatically compute API signatures consisting of request and response fields. We evaluated the framework through a qualitative survey of nine mainframe developers with an average of 15 years of experience, using the public GENAPP application and two industrial mainframe applications. The results show that the framework identifies additional candidate APIs and reduces implementation effort for APIfication. The API-signature computation has been incorporated into IBM watsonx Code Assistant for Z Refactoring Assistant. We further validated the identified APIs by executing them on an IBM Z mainframe system, demonstrating the practical viability of the approach.

cs.SE↗