{ "format": "agent-article-v0", "article_id": "EA-126", "canonical_url": "/tools/ea-126-planner-executor", "page_type": "reference", "claim_ids": [], "evidence_ids": [ "EVD-EA-126-URL-310E25743A23", "EVD-EA-126-URL-7647307FE498", "EA-126-R11-RESEARCH", "EA-126-R11-WEB-READBACK" ], "unknown_ids": [ "EA-126-UNKNOWN-01", "EA-126-UNKNOWN-CURRENTNESS" ], "operational_object_ids": [], "provider": [], "models": [], "plans": [], "surfaces": [], "regions": [], "current_status": "UNKNOWN", "observation_status": "SPEC_ONLY", "freshness": "UNKNOWN", "last_verified": null, "supersedes": [], "superseded_by": [], "provenance": [ "COMM-042-EA0126", "SRC-EA-126-URL-310E25743A23", "SRC-EA-126-URL-7647307FE498", "WEBBODY-EA0126-R8-C59F9E49" ], "article_sources": [ { "id": "EVD-EA-126-URL-310E25743A23", "evidence_class": "RESEARCH", "title": "source class: primary research paper; exact title: Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models; author/org: Lei Wang et al.; c", "href": "https://arxiv.org/abs/2305.04091", "supports": [], "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。", "origin_class": "SOURCE_PACKET_REFERENCE", "source_or_experiment_ids": [ "SRC-EA-126-URL-310E25743A23" ], "supporting_locators": [ "EVIDENCE TABLE", "- source class: primary research paper; exact title: Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models; author/org: Lei Wang et al.; canonical URL: https://arxiv.org/abs/2305.04091" ], "scope": "EVIDENCE TABLE", "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED" }, { "id": "EVD-EA-126-URL-7647307FE498", "evidence_class": "RESEARCH", "title": "source class: primary research paper; exact title: ReAct: Synergizing Reasoning and Acting in Language Models; author/org: Shunyu Yao et al.; canonical URL:", "href": "https://arxiv.org/abs/2210.03629", "supports": [], "limit": "元の監査済み資料に記載された参照先です。URLの復元は、資料の最新性や全主張の支持を新たに検証したことを意味しません。", "origin_class": "SOURCE_PACKET_REFERENCE", "source_or_experiment_ids": [ "SRC-EA-126-URL-7647307FE498" ], "supporting_locators": [ "EVIDENCE TABLE", "- source class: primary research paper; exact title: ReAct: Synergizing Reasoning and Acting in Language Models; author/org: Shunyu Yao et al.; canonical URL: https://arxiv.org/abs/2210.03629" ], "scope": "EVIDENCE TABLE", "provenance_status": "SOURCE_PACKET_REFERENCE_PROJECTED" }, { "id": "EA-126-R11-RESEARCH", "evidence_class": "RESEARCH", "title": "EA-126 監査済みresearch / canonical revision binding", "supports": [ "Plan-and-Solveは、全体をsubtaskへ分けるplanと、planに従うsubtask executionの二成分を明示し、GPT-3の10 datasets/3 reasoning problemsでZero-shot-CoT等との比較を報告する。これはplan→executeという制御形の研究根拠だが、planner/executorを異なるmodel/providerにすること、executorが外部transactionを安全に行えること、planが常に完全であることは確立しない。ReActはreasoningとactingをinterleaveし、外部観測でplanを更新するため、固定planと反応型planのトレードオフを示すcounter-contextである。" ], "limit": "canonical監査・revision bindingの存在を示す。外部ソースcurrentnessが本文またはcontrol-planeで確定していない場合は、currentnessをUNKNOWNのまま保持する。", "origin_class": "AUDITED_CANONICAL_RESEARCH_PACKET", "source_or_experiment_ids": [ "COMM-042-EA0126" ], "supporting_locators": [ "COMM-042-EA0126", "EA-126-R5-LEGACY-MIGRATION-20261008-R1", "content/staging/EA-126.md" ], "scope": "監査済みEA本文・canonical revision・既存Source Graph relationの範囲。欠損currentnessは推測で閉じない。", "provenance_status": "AUDIT_PASS_CANONICAL_BINDING" }, { "id": "EA-126-R11-WEB-READBACK", "evidence_class": "IMPL", "title": "EA-126 Web body identity / canonical readback", "supports": [ "canonical staging bodyのSHA-256一致とR9 canonical ledger readbackをProduction projectionの入力として検証する。" ], "limit": "body identity/readbackは本文claim自体の真偽や外部ソースの最新性を追加保証しない。", "checked_at": null, "origin_class": "IMPLEMENTATION_READBACK", "source_or_experiment_ids": [ "WEBBODY-EA0126-R8-C59F9E49" ], "supporting_locators": [ "WEBBODY-EA0126-R8-C59F9E49", "content/staging/EA-126.md", "EA-126-R5-LEGACY-MIGRATION-20261008-R1" ], "scope": "R9/R10 canonical readback後のProduction生成時body identity。", "provenance_status": "CANONICAL_READBACK_BODY_HASH_MATCH" } ], "research_sections": [ { "heading": "WHY_THIS_MATTERS_TO_UPDATE", "text": "強い推論モデルを計画、安価/決定論的系を実行に割り当てる構成は費用と権限を分離し得るが、planの不完全性とexecutorの過剰権限が同時に危険になる。" }, { "heading": "SCOPE_ROUTING", "text": "multi-LLM / implementation / cost / security / evaluation" }, { "heading": "CLAIM_OR_RESEARCH_QUESTION", "text": "Plan-and-Solve原著が示すplan→subtask executionは何を確立するか。plannerとexecutorを異種modelに分けること、または計画を信頼して実行することまで確立するか。" }, { "heading": "EVIDENCE TABLE", "text": "- source class: primary research paper; exact title: Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models; author/org: Lei Wang et al.; canonical URL: https://arxiv.org/abs/2305.04091; publication/update/access: submitted 2023-05-06, revised v3 2023-05-26 / accessed 2026-09-13 JST; version/release/tag/commit/API revision/model ID: arXiv v3 / ACL 2023; supporting locator: abstract lines 8-24（two components: devise plan to divide task into subtasks, then carry out subtasks; ten datasets/three reasoning problems; GPT-3 results）; evidence label: RESEARCH; relation: support / limit.\n- source class: primary research paper; exact title: ReAct: Synergizing Reasoning and Acting in Language Models; author/org: Shunyu Yao et al.; canonical URL: https://arxiv.org/abs/2210.03629; publication/update/access: revised v3 2023-03-10 / accessed 2026-09-13 JST; version/release/tag/commit/API revision/model ID: arXiv v3; supporting locator: abstract lines 16-21（reasoning traces and actions interleaved, plans updated by external observations, HotpotQA/FEVER/ALFWorld/WebShop scope）; evidence label: RESEARCH; relation: counter / context." }, { "heading": "FINDINGS", "text": "Plan-and-Solveは、全体をsubtaskへ分けるplanと、planに従うsubtask executionの二成分を明示し、GPT-3の10 datasets/3 reasoning problemsでZero-shot-CoT等との比較を報告する。これはplan→executeという制御形の研究根拠だが、planner/executorを異なるmodel/providerにすること、executorが外部transactionを安全に行えること、planが常に完全であることは確立しない。ReActはreasoningとactingをinterleaveし、外部観測でplanを更新するため、固定planと反応型planのトレードオフを示すcounter-contextである。" }, { "heading": "LIMITS / COUNTEREVIDENCE / FAILURE ENVELOPE", "text": "plan omission、stale world state、executor interpretation drift、subtask dependency、error propagation、tool permission overreach。planを一度承認した後の環境変化・権限変更・外部副作用は別監視が必要。" }, { "heading": "AVAILABILITY_GATES", "text": "planner/executor model/API/version、tool permission、plan schema、state snapshot、approval、budget、rollback、provider data policyを確認する。実model混成・実tool executorは未実施。" }, { "heading": "IMPLEMENTATION_STATE", "text": "plannerはtyped plan/dry-runのみ、executorはallowlisted deterministic/tool actions、state readback、step-level approval、replan on driftを設計候補とする。" }, { "heading": "BENCHMARK_OR_EVAL", "text": "plan completeness、dependency correctness、executor success、state divergence、action safety、replan frequency、latency/cost、human approval burdenを別scoreにする。計画品質を実行安全性に代用しない。" }, { "heading": "SECURITY / PRIVACY / AUDIT / COST / OPERABILITY", "text": "SECURITY/PRIVACY: plannerへ秘密を渡さず、executorに最小権限。AUDIT: plan version/hash、model IDs、state snapshot、step requests/results、approvals、replan。COST: planner expensive model、executor calls、readback。OPERABILITY: max steps、stop on drift、compensation/revert、dead-letter。" }, { "heading": "SAFE_WORDING_FOR_FUTURE_BOOK", "text": "「Plan-and-Solve研究はplan→subtask executionの構造と特定推論taskでの比較結果を示す。Planner–Executor異種化は費用・権限分離の設計候補であり、planの完全性、executorの安全性、外部actionの可逆性を自動保証しない。ReActのようなinterleavingも代替設計である。」" }, { "heading": "UPDATE_WORK_HANDOFF", "text": "本文ではplanner/executorのmodel、protocol、state、tool permission、approval、replan、rollback、costを明記し、GPT-3/ten-dataset研究値の適用範囲を固定する。" }, { "heading": "OPEN_UNKNOWN", "text": "planner/executor異種化の因果効果、plan drift検知、再計画条件、deterministic executorの実現、長期taskでの費用/安全性。" }, { "heading": "STOP_CONDITION", "text": "plan生成成功やexecutorの一step成功を、全体完了・安全・正しさの証拠としない。\nUNIT_CLOSE_STATUS: INDIVIDUAL_BOUNDED_CLOSE / EVIDENCE_RECORDED / UNKNOWN_PRESERVED\n\n---" } ], "human_url": "/tools/ea-126-planner-executor", "markdown_url": "/agent/article.md?path=%2Ftools%2Fea-126-planner-executor", "json_url": "/agent/article.json?path=%2Ftools%2Fea-126-planner-executor" }