{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.13715", "kind": "paper", "url": "https://arxiv.org/abs/2606.13715"}], "categories": ["Agents", "General AI", "Language & Knowledge"], "description": {"en": "WorkBench evaluates workplace agents by task completion and unintended harmful actions."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.13715", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.13715", "title": "WorkBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2606.13715", "modality": null, "name": "WorkBench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-09-25T04:40:22Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_workbench_686f5260", "origin_source": "arxiv", "recognition_confidence": "0.6", "record_sha256": "fccb6f147cf530deb11e689ecaea05de072cf31997e312f91f28127cb4e941b6", "relation": "extends", "review_model": "codex-cli-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T18:27:27.771141Z", "source_url": "https://arxiv.org/abs/2606.13715"}, "publisher": null, "released": "2026-06-10", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.13715", "source_url": "https://arxiv.org/abs/2606.13715"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-13715", "source": "claire_radar", "source_benchmark_id": "2606.13715", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The near-saturation result and longitudinal capability-safety findings may attract attention in agent-evaluation discussions.", "score": 76}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T18:27:27.771141Z", "inputHash": "cc0ac743f31d786b48f3ac580d73c0f2ef606ed54f016479f942f226a8bc7af2", "model": "codex-cli-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "WorkBench Revisited: Workplace Agents Two Years On", "canonicalNameSource": "paper_title", "decisionReason": "The source names a benchmark, reports comparable task-completion and harm measurements, and explicitly releases an updated version with data and code improvements.", "evaluationEvidence": {"reuse": "We release an updated version of the benchmark with data and code quality improvements", "scoring": "now completes 98%", "task": "completed just 43% of tasks"}, "model": "codex-cli-default", "reviewedAt": "2026-09-24T18:27:27.771141Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "WorkBench evaluates workplace agents by task completion and unintended harmful actions.", "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"], "snippet": "We release an updated version of the benchmark with data and code quality improvements, new model scores, and analysis of agent progress on WorkBench since 2024."}, "familyId": "bmf_d85b40105dd0", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_workbench_686f5260", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2606.13715", "pdf": "https://arxiv.org/pdf/2606.13715", "project": null, "report": "https://arxiv.org/abs/2606.13715"}, "metrics": [], "motivation": "The best agent on WorkBench in March 2024, GPT-4, completed just 43% of tasks.", "name": "WorkBench", "oneLine": "WorkBench evaluates workplace agents by task completion and unintended harmful actions.", "paperTitle": "WorkBench Revisited: Workplace Agents Two Years On", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.13715", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Paper only", "recognitionConfidence": 0.6, "relation": "extends", "releasedAt": "2026-06-10", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Olly Styles", "Sam Miller"], "categories": ["cs.AI", "cs.CL", "cs.MA"], "id": "2606.13715", "title": "WorkBench Revisited: Workplace Agents Two Years On", "type": "arxiv", "url": "https://arxiv.org/abs/2606.13715"}, "sourceUpdatedAt": "2026-07-01T14:36:06Z", "topics": ["Agents"], "venueAttempts": [], "whyItMatters": "The updated release supports comparison of workplace-agent capability, safety, and cost over time."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
