{"record": {"aliases": ["One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows"], "artifacts": [{"id": "arxiv:2608.19741", "kind": "paper", "url": "https://arxiv.org/abs/2608.19741"}, {"id": "gh:microsoft/thinkingbox", "kind": "repo", "url": "https://github.com/microsoft/thinkingbox"}], "authors": ["Zhuochun Li", "Youngmin Ko", "Ali Keramati", "Tuhin Kundu", "Liang-Chun Tsai", "Nicola Ferri", "Mirco Milletari", "Jiaxiang Liu", "Susana Palmaz Lopez Pelaez", "Yuepeng Wang", "Vadim Smolyakov", "Xiang Jiang", "Kjartan Olafsson", "Tommy Guy"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.19741", "source_url": "https://arxiv.org/abs/2608.19741"}, "categories": ["General AI", "Language & Knowledge", "cs.CL"], "description": {"en": "Evaluates LLM agents on 507 stateful business workflows in an executable sandbox, with task-specific checks that accept valid trajectories and reject wrong, missing, or extra effects."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.19741", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.19741", "title": "Thinkingbox-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["University of Pittsburgh", "Northwestern University", "University of California, Irvine", "Microsoft (United States)"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2608.19741", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2608.19741", "source_url": "https://openalex.org/W7203928100"}, "key": "claire-radar:2608.19741", "modality": null, "name": "Thinkingbox-Bench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_one-success-isn-t-reliability-thinkingbox-_a377202a", "origin_source": "arxiv", "recognition_confidence": "0.7", "record_sha256": "c0b37f48b075d64739e13db3f5c7e1caf8b8c53634bc2f0bb1070b9d121e739e", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-08-26T06:15:55.708965Z", "source_url": "https://arxiv.org/abs/2608.19741"}, "publisher": null, "released": "2026-08-20", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.19741", "source_url": "https://arxiv.org/abs/2608.19741"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-19741", "source": "claire_radar", "source_benchmark_id": "2608.19741", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-08-26", "attemptedAt": "2026-08-25T23:36:26.005677Z", "githubRepo": "https://github.com/microsoft/thinkingbox", "githubScope": "benchmark_repo", "githubStars": 20, "hfDailySubmittedAt": "2026-08-25T00:00:00.000Z", "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 9, "hfPaperUrl": "https://huggingface.co/papers/2608.19741", "observedAt": "2026-08-25T23:36:26.005677Z", "signalStatus": {"githubStars": {"observedAt": "2026-08-25T23:36:26.005677Z", "state": "fresh"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"observedAt": "2026-08-25T23:36:26.005677Z", "state": "fresh"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The focus on stateful business workflows and agent reliability, combined with Microsoft affiliation and open-source release, likely drives above-average immediate interest.", "score": 65}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-26T06:15:55.708965Z", "inputHash": "68bd8f809edbe512ec070ff8dd5c435d086e10f13fe3cbada5b11b243391e171", "model": "deepseek-v4-pro"}, "curation": {"canonicalNameEvidence": "Thinkingbox-bench contains 507 policy-conditioned workflows across numerous scenarios", "canonicalNameSource": "abstract", "decisionReason": "The paper introduces a named benchmark with a sandbox framework, explicit evaluation checks, and release of code and data via GitHub, meeting the criteria for a reusable benchmark.", "model": "deepseek-v4-pro", "reviewedAt": "2026-08-26T06:15:55.708965Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates LLM agents on 507 stateful business workflows in an executable sandbox, with task-specific checks that accept valid trajectories and reject wrong, missing, or extra effects.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "We release both Thinkingbox and Thinkingbox-Bench: https://github.com/microsoft/thinkingbox"}, "familyId": "bmf_c4ef2d2a7c2c", "firstSeenAt": "2026-08-21", "heat": null, "id": "bm_one-success-isn-t-reliability-thinkingbox-_a377202a", "indexedAt": "2026-08-21T04:30:09.569171Z", "industrySectors": [], "links": {"code": "https://github.com/microsoft/thinkingbox", "data": null, "paper": "https://arxiv.org/abs/2608.19741", "pdf": "https://arxiv.org/pdf/2608.19741", "project": null, "report": "https://arxiv.org/abs/2608.19741"}, "metrics": [], "motivation": "Recent agent benchmarks increasingly ground evaluation in executable environments, from code repair to web navigation, app APIs, and function calling.", "name": "Thinkingbox-Bench", "oneLine": "Evaluates LLM agents on 507 stateful business workflows in an executable sandbox, with task-specific checks that accept valid trajectories and reject wrong, missing, or extra effects.", "paperTitle": "One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2608.19741", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-21T04:30:09.569171Z"}, "publications": [], "publishers": [{"name": "Microsoft", "organizationType": "company-research-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/microsoft/thinkingbox"}], "ranking": {"30d": {"confidence": "High", "coverage": 0.85, "growth": {"components": {"githubStars": {"percentile": null, "value": null}, "hfDatasetDownloads": {"percentile": null, "value": null}, "hfPaperUpvotes": {"percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 30}, "level": {"components": {"githubStars": {"percentile": 0.8464285714285714, "value": 20}, "hfDatasetDownloads": {"percentile": null, "value": null}, "hfPaperUpvotes": {"percentile": 0.5972222222222222, "value": 9}}, "confidence": "High", "coverage": 0.85, "rank": 27, "score": 72}, "method": "attention-ranking-v11", "rank": 27, "score": 72}, "90d": {"confidence": "Medium", "coverage": 0.7, "growth": {"components": {"githubStars": {"percentile": null, "value": null}, "hfDatasetDownloads": {"percentile": null, "value": null}, "hfPaperUpvotes": {"percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 90}, "level": {"components": {"githubStars": {"percentile": 0.7425068119891008, "value": 20}, "hfDatasetDownloads": {"percentile": null, "value": null}, "hfPaperUpvotes": {"percentile": 0.7116279069767442, "value": 9}}, "confidence": "Medium", "coverage": 0.7, "rank": 94, "score": 67}, "method": "attention-ranking-v11", "rank": 94, "score": 67}}, "readiness": "Runnable", "recognitionConfidence": 0.7, "relation": "introduces", "releasedAt": "2026-08-20", "source": {"authors": ["Zhuochun Li", "Youngmin Ko", "Ali Keramati", "Nicola Ferri", "Susana Palmaz Lopez Pelaez", "Liang-Chun Tsai", "Calvin Wang", "Mirco Milletari", "Tuhin Kundu", "Vadim Smolyakov", "Kjartan Olafsson", "Tommy Guy"], "categories": ["cs.CL", "cs.DB"], "id": "2608.19741", "title": "One Success Isn't Reliability: Thinkingbox, a Sandbox and Benchmark for Agents in Stateful Business Workflows", "type": "arxiv", "url": "https://arxiv.org/abs/2608.19741"}, "sourceUpdatedAt": "2026-08-20T07:37:57Z", "topics": ["cs.CL"], "venueAttempts": [], "whyItMatters": "Moves beyond response-level or tool-call-level evaluation by requiring correct persistent state transitions, addressing a key gap in assessing agents for consequential business tasks."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
