Skip to content
Log in
Preprint

One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows

Zhuochun Li Youngmin Ko Ali Keramati Nicola Ferri Susana Palmaz Lopez Pelaez Liang-Chun Tsai Calvin Wang Mirco Milletari Tuhin Kundu V. Smolyakov Kjartan Ólafsson Tommy Guy
Aug 2026 · 0 citations · 44 references
Computer Science

TL;DR

Thinkingbox is introduced, a sandbox for tool-agent-user interaction that provides isolated MCP-compatible tool sessions, complete execution traces, and outcome evaluation over terminal backend state, and Thinkingbox-bench reveals a large gap between occasionally finding a successful trajectory and reliably completing stateful business tasks.

Abstract

Recent agent benchmarks increasingly ground evaluation in executable environments, from code repair to web navigation, app APIs, and function calling. Yet completing consequential work beyond code requires more than producing a plausible response or valid tool call: agents must gather missing information over multiple turns, follow domain policies, coordinate dependent tools, and realize the correct persistent state transition without collateral effects. In this paper, we introduce Thinkingbox, a sandbox for tool-agent-user interaction that provides isolated MCP-compatible tool sessions, complete execution traces, and outcome evaluation over terminal backend state. Built on this sandbox, Thinkingbox-bench contains 507 policy-conditioned workflows across numerous scenarios, including retail, hospitality, auto insurance, neobank internal IT, and consulting IT/HR support. Each attempt is evaluated by task-specific executable checks that accept valid trajectories while rejecting wrong, missing, or extra effects; designated tasks additionally check required properties of the final response. Across proprietary and open-weight models, the strongest model Claude Opus 5 achieves 66.50% pass@1, but only 47.53% pass^20. Moreover, many failed trials show clean termination and valid state-changing actions, showing that response or tool-call-level signals are not clear proxies for end-to-end task completion. Thinkingbox-bench reveals a large gap between occasionally finding a successful trajectory and reliably completing stateful business tasks. We release both Thinkingbox and Thinkingbox-Bench: https://github.com/microsoft/thinkingbox

DOI arXiv Semantic Scholar
View source

Similar papers

Preprint Jul 2026

E-Bench: Benchmarking Multi-Step Tool-Use Agents in Real-World Product Scenarios

E-Bench is introduced, a fully synthetic benchmark with 323 state-changing tasks across three product domains: Honor of Kings, QQ Music, and Tencent Meeting, and it shows that multi-step tool use remains challenging: Pass^3 stays below 60% for the strongest models, and even with code execution in the E-Bench-Code extension, reliability remains below 70%.

Weihuang Zheng, Tianyuan Zou, Eileen Ye et al. · 1 citation
Open access Jul 2026

TraceUX: An Explainable Rule-Based Framework for Context-Aware Static UX Evaluation

User experience (UX) evaluation is central to software quality, yet it remains difficult to integrate into software engineering workflows in a systematic, explainable, and early-stage manner. This paper presents TraceUX, a framework for operationalizing UX heuristics and design guidance into a rule-based static evaluation pipeline that combines machine-interpretable formalization, executability-aware assessment, context-sensitive scoring, and actionable reporting. The framework is instantiated using Apple Human Interface Guidelines, Swift abstract syntax trees, and mobile games, and implemented in a proof-of-concept tool named TraceHIG. Evaluation was conducted in four layers: analysis of the full rule repository, controlled synthetic validation with injected violations, baseline assessment of 12 public Swift game projects, and a case study on one project. The full repository contained 206 rules; after excluding non-iOS yet platform-specific rules, 193 rules were retained for the downstream experiments. In controlled validation, 216 injected violations yielded 99.2% precision, 61.6% recall, and an F1-score of 0.760. In baseline analysis, overall project scores ranged from 41.6 to 88.0, reflecting rule-conformance spread under the instantiated rule base rather than direct measures of UX quality. The case study demonstrated that profile-aware scoring can yield materially different UX assessments for the same codebase under different game configurations, highlighting the importance of app profiling in static UX evaluation. These findings show that a meaningful subset of UX knowledge can be operationalized into explainable, context-aware static analysis that provides structured and actionable decision support while complementing, rather than replacing, manual and empirical UX evaluation.

Fouzia Alzhrani · 0 citations
Preprint Jul 2026

SLBench: Evaluating How LLM Agents Follow Logical Relations in Skills

This work introduces SkillLogic, a framework for analyzing logical relations in skill files and constructing executable tests from them, and establishes logical-relation following as a distinct reliability challenge for skill-guided agents.

Xuan Chen, Chengpeng Wang, Lu Yan et al. · 0 citations
Preprint Jul 2026

StructureClaw: Traceable LLM Agents and an Executable Benchmark for Structural Engineering Workflows

This work presents StructureClaw, an artifact-centered workbench in which LLM agents operate through governed engineering skills, typed tools, shared artifact state, and local analysis backends, together with StructureClaw-Bench, an executable benchmark of 150 controlled scenarios spanning standard workflows, interactive robustness, and multimodal structural-model reconstruction.

Sizhong Qin, Yi Gu, Yao Jiang et al. · 0 citations