{"id":"14f66bbfc639af860d443396c291d8906dae763f3ebd672785d390976d5f3986","title":"Jaewon Chu and coauthors propose AgentGrad for multi-agent systems","argument":"Jaewon Chu and coauthors propose AgentGrad, a framework for optimizing prompts in multi-agent systems (MAS), where specialized agents work together and performance depends on each agent’s prompt. They argue that existing textual gradient methods may update a prompt without checking whether the change fixes a failure, and may combine gradients from unrelated failures. AgentGrad uses sequential intervention: it adds a hint to one agent at a time, then uses the modified output of an agent whose change resolves the failure to supervise gradient extraction. Semantic textual gradient abstraction groups similar gradients into generalized gradients that capture shared corrective patterns. The authors report state-of-the-art performance across five MAS benchmarks, with 2.5 times less wall-clock optimization time and 21.8% lower optimization cost on average than the next-best baselines. The approach ties prompt updates to the agent associated with a failure and groups similar corrections; whether that improves performance beyond the reported benchmarks is not established by the abstract. The paper was accepted to NeurIPS 2026.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The evidence is the arXiv abstract; it does not provide benchmark-level results or details of the experimental measurements. The reported performance and optimization savings are the authors’ claims about their experiments; the abstract does not establish real-world reliability or performance beyond the five MAS benchmarks.","speaker":{"name":"Chu, Jaewon, Seo, Jinwoo, Cho, Jaewon, Na, Jeehye, Xiong, Yunyang, Kim, Youngdae, Kim, Hyunwoo J.","roleAtSource":"Article author"},"source":{"title":"AgentGrad: Intervention-guided Prompt Optimization for Multi-Agent Systems","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2609.08572","publishedAt":"2026-09-08","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors identify prompt selection without verifying its effect and lack of agent-level supervision as limitations in gradient extraction.","excerpt":"previous works select a prompt to update without verifying whether modifying it resolves the failure, and derive gradients without agent-level supervision","url":"https://arxiv.org/abs/2609.08572","relationship":"Source-stated"},{"text":"AgentGrad uses sequential intervention to identify which agent’s modification resolves a failure, then uses that agent’s modified output as supervision.","excerpt":"sequential intervention injects a hint into one agent at a time to identify the target agent whose modification resolves the failure","url":"https://arxiv.org/abs/2609.08572","relationship":"Source-stated"},{"text":"AgentGrad clusters semantically similar gradients and abstracts each cluster into a generalized gradient intended to capture shared corrective patterns.","excerpt":"clusters semantically similar gradients to prevent mixing unrelated failure modes, and abstracts each cluster into a generalized gradient","url":"https://arxiv.org/abs/2609.08572","relationship":"Source-stated"},{"text":"The authors report state-of-the-art performance on five MAS benchmarks and average reductions in optimization time and cost compared with the next-best baselines.","excerpt":"achieves state-of-the-art performance across five MAS benchmarks while reducing wall-clock optimization time by 2.5 times and optimization cost by 21.8% on average","url":"https://arxiv.org/abs/2609.08572","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2609.08572","checkedAt":"2026-11-10T04:50:09.070Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How do AgentGrad’s results vary across the five MAS benchmarks, and what evidence shows its prompts generalize beyond them?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Jaewon Chu and coauthors propose AgentGrad for multi-agent systems","summary":"Jaewon Chu and coauthors propose AgentGrad, a framework for optimizing prompts in multi-agent systems (MAS), where specialized agents work together and performance depends on each agent’s prompt. They argue that existing textual gradient methods may update a prompt without checking whether the change fixes a failure, and may combine gradients from unrelated failures. AgentGrad uses sequential intervention: it adds a hint to one agent at a time, then uses the modified output of an agent whose change resolves the failure to supervise gradient extraction. Semantic textual gradient abstraction groups similar gradients into generalized gradients that capture shared corrective patterns. The authors report state-of-the-art performance across five MAS benchmarks, with 2.5 times less wall-clock optimization time and 21.8% lower optimization cost on average than the next-best baselines. The approach ties prompt updates to the agent associated with a failure and groups similar corrections; whether that improves performance beyond the reported benchmarks is not established by the abstract. The paper was accepted to NeurIPS 2026.","question":"How do AgentGrad’s results vary across the five MAS benchmarks, and what evidence shows its prompts generalize beyond them?","claims":[{"text":"The authors identify prompt selection without verifying its effect and lack of agent-level supervision as limitations in gradient extraction.","excerpt":"previous works select a prompt to update without verifying whether modifying it resolves the failure, and derive gradients without agent-level supervision","support":"source-stated"},{"text":"AgentGrad uses sequential intervention to identify which agent’s modification resolves a failure, then uses that agent’s modified output as supervision.","excerpt":"sequential intervention injects a hint into one agent at a time to identify the target agent whose modification resolves the failure","support":"source-stated"},{"text":"AgentGrad clusters semantically similar gradients and abstracts each cluster into a generalized gradient intended to capture shared corrective patterns.","excerpt":"clusters semantically similar gradients to prevent mixing unrelated failure modes, and abstracts each cluster into a generalized gradient","support":"source-stated"},{"text":"The authors report state-of-the-art performance on five MAS benchmarks and average reductions in optimization time and cost compared with the next-best baselines.","excerpt":"achieves state-of-the-art performance across five MAS benchmarks while reducing wall-clock optimization time by 2.5 times and optimization cost by 21.8% on average","support":"source-stated"}],"limitations":["The evidence is the arXiv abstract; it does not provide benchmark-level results or details of the experimental measurements.","The reported performance and optimization savings are the authors’ claims about their experiments; the abstract does not establish real-world reliability or performance beyond the five MAS benchmarks."],"mediaId":"e121ff84f01d21f9fde526eb56f160dba769b549384e3795973328a4d0c6a81a","correctionReason":null,"id":"14f66bbfc639af860d443396c291d8906dae763f3ebd672785d390976d5f3986","publicationId":"e31cf00b-c446-425c-8979-e72749dcadaf","version":1,"sourceUrl":"https://arxiv.org/abs/2609.08572","sourceTitle":"AgentGrad: Intervention-guided Prompt Optimization for Multi-Agent Systems","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-09-08","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-09T05:03:39.821Z","createdAt":"2026-11-10T03:46:31.867Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"e121ff84f01d21f9fde526eb56f160dba769b549384e3795973328a4d0c6a81a","kind":"article","url":"https://arxiv.org/abs/2609.08572","sourceUrl":"https://arxiv.org/abs/2609.08572","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-10T03:16:37.161Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2609.08572","checkedAt":"2026-11-10T04:50:09.070Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"e121ff84f01d21f9fde526eb56f160dba769b549384e3795973328a4d0c6a81a","confidence":0.87,"provider":"typesafe","model":"jev-1.13.0","inputHash":"81be96dd3dadd128baf924359ebcd7d9bd590c967f372b38cd69306d60bed647","decidedAt":"2026-11-10T03:16:37.964Z"},"provenance":{"textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observationId":"48cb643edddb29d63eab6f4c60250d8b61e329b366102b57ff127e77f4f34441","retrievedAt":"2026-11-10T03:16:37.039Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"b8cd52b7-0ff5-4ae7-80fa-571ebeed65a7"},"reviewer":{"model":"gpt-6-luna","runId":"b7a7c37f-0a94-41f3-8d21-a27022a1f877"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-10T03:46:31.867Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Chu, Jaewon","url":null,"kind":"unknown"},{"name":"Seo, Jinwoo","url":null,"kind":"unknown"},{"name":"Cho, Jaewon","url":null,"kind":"unknown"},{"name":"Na, Jeehye","url":null,"kind":"unknown"},{"name":"Xiong, Yunyang","url":null,"kind":"unknown"},{"name":"Kim, Youngdae","url":null,"kind":"unknown"},{"name":"Kim, Hyunwoo J.","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Chu, Jaewon","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Chu, Jaewon","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2609.08572","locator":"head/meta[name=author|citation_author]"},"assertionId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T04:50:09.070Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36","sourceUrl":"https://arxiv.org/abs/2609.08572","contentSha256":"8fba1ec8f95099e9041e160de6dd94a45ea8e4a619901564934097be61fafee5","textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observedAt":"2026-11-10T04:50:09.070Z"}},{"actor":{"name":"Seo, Jinwoo","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Seo, Jinwoo","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2609.08572","locator":"head/meta[name=author|citation_author]"},"assertionId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T04:50:09.070Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36","sourceUrl":"https://arxiv.org/abs/2609.08572","contentSha256":"8fba1ec8f95099e9041e160de6dd94a45ea8e4a619901564934097be61fafee5","textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observedAt":"2026-11-10T04:50:09.070Z"}},{"actor":{"name":"Cho, Jaewon","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Cho, Jaewon","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2609.08572","locator":"head/meta[name=author|citation_author]"},"assertionId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T04:50:09.070Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36","sourceUrl":"https://arxiv.org/abs/2609.08572","contentSha256":"8fba1ec8f95099e9041e160de6dd94a45ea8e4a619901564934097be61fafee5","textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observedAt":"2026-11-10T04:50:09.070Z"}},{"actor":{"name":"Na, Jeehye","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Na, Jeehye","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2609.08572","locator":"head/meta[name=author|citation_author]"},"assertionId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T04:50:09.070Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36","sourceUrl":"https://arxiv.org/abs/2609.08572","contentSha256":"8fba1ec8f95099e9041e160de6dd94a45ea8e4a619901564934097be61fafee5","textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observedAt":"2026-11-10T04:50:09.070Z"}},{"actor":{"name":"Xiong, Yunyang","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Xiong, Yunyang","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2609.08572","locator":"head/meta[name=author|citation_author]"},"assertionId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T04:50:09.070Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36","sourceUrl":"https://arxiv.org/abs/2609.08572","contentSha256":"8fba1ec8f95099e9041e160de6dd94a45ea8e4a619901564934097be61fafee5","textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observedAt":"2026-11-10T04:50:09.070Z"}},{"actor":{"name":"Kim, Youngdae","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Kim, Youngdae","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2609.08572","locator":"head/meta[name=author|citation_author]"},"assertionId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T04:50:09.070Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36","sourceUrl":"https://arxiv.org/abs/2609.08572","contentSha256":"8fba1ec8f95099e9041e160de6dd94a45ea8e4a619901564934097be61fafee5","textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observedAt":"2026-11-10T04:50:09.070Z"}},{"actor":{"name":"Kim, Hyunwoo J.","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Kim, Hyunwoo J.","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2609.08572","locator":"head/meta[name=author|citation_author]"},"assertionId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36:participation:6","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T04:50:09.070Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36","sourceUrl":"https://arxiv.org/abs/2609.08572","contentSha256":"8fba1ec8f95099e9041e160de6dd94a45ea8e4a619901564934097be61fafee5","textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observedAt":"2026-11-10T04:50:09.070Z"}}],"participationCoverage":{"total":7,"returned":7,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2609.08572","snapshotId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36","contentSha256":"8fba1ec8f95099e9041e160de6dd94a45ea8e4a619901564934097be61fafee5","textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observedAt":"2026-11-10T04:50:09.070Z","basis":"publisher-page-metadata","sourceLinks":49,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/14f66bbfc639af860d443396c291d8906dae763f3ebd672785d390976d5f3986/source"}},"canonicalUrl":"https://escapemeridian.com/readings/jaewon-chu-and-coauthors-propose-agentgrad-for-multi-agent-14f66bbf","formats":{"html":"/readings/jaewon-chu-and-coauthors-propose-agentgrad-for-multi-agent-14f66bbf","markdown":"/readings/jaewon-chu-and-coauthors-propose-agentgrad-for-multi-agent-14f66bbf.md","json":"/readings/jaewon-chu-and-coauthors-propose-agentgrad-for-multi-agent-14f66bbf.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[{"id":"term-benchmark","name":"Benchmark","explanation":"A defined set of tests used to compare systems. Results depend on the tasks and measures selected.","referenceUrl":"https://arxiv.org/abs/2211.09110"}],"entities":[{"id":"arxiv","name":"arXiv","kind":"organization","context":"arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.","referenceUrls":["https://arxiv.org/","https://info.arxiv.org/help/api/index.html"],"path":"/entities/arxiv","json":"/entities/arxiv.json","markdown":"/entities/arxiv.md"},{"id":"time","name":"TIME","kind":"organization","context":"TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.","referenceUrls":["https://time.com/"],"path":"/entities/time","json":"/entities/time.json","markdown":"/entities/time.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"14f66bbfc639af860d443396c291d8906dae763f3ebd672785d390976d5f3986","revision":"e31cf00b-c446-425c-8979-e72749dcadaf","revisionKind":"publication-id","title":"Jaewon Chu and coauthors propose AgentGrad for multi-agent systems","url":"https://escapemeridian.com/readings/jaewon-chu-and-coauthors-propose-agentgrad-for-multi-agent-14f66bbf"},"source":{"type":"source","id":"https://arxiv.org/abs/2609.08572","revision":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","revisionKind":"sha256-source-text","title":"AgentGrad: Intervention-guided Prompt Optimization for Multi-Agent Systems","url":"https://arxiv.org/abs/2609.08572"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter."},"revisions":[{"id":"14f66bbfc639af860d443396c291d8906dae763f3ebd672785d390976d5f3986","title":"Jaewon Chu and coauthors propose AgentGrad for multi-agent systems","argument":"Jaewon Chu and coauthors propose AgentGrad, a framework for optimizing prompts in multi-agent systems (MAS), where specialized agents work together and performance depends on each agent’s prompt. They argue that existing textual gradient methods may update a prompt without checking whether the change fixes a failure, and may combine gradients from unrelated failures. AgentGrad uses sequential intervention: it adds a hint to one agent at a time, then uses the modified output of an agent whose change resolves the failure to supervise gradient extraction. Semantic textual gradient abstraction groups similar gradients into generalized gradients that capture shared corrective patterns. The authors report state-of-the-art performance across five MAS benchmarks, with 2.5 times less wall-clock optimization time and 21.8% lower optimization cost on average than the next-best baselines. The approach ties prompt updates to the agent associated with a failure and groups similar corrections; whether that improves performance beyond the reported benchmarks is not established by the abstract. The paper was accepted to NeurIPS 2026.","topicIds":["intelligence-and-discovery"],"domain":"science","claimType":"attributed reading","limits":"The evidence is the arXiv abstract; it does not provide benchmark-level results or details of the experimental measurements. The reported performance and optimization savings are the authors’ claims about their experiments; the abstract does not establish real-world reliability or performance beyond the five MAS benchmarks.","speaker":{"name":"Chu, Jaewon, Seo, Jinwoo, Cho, Jaewon, Na, Jeehye, Xiong, Yunyang, Kim, Youngdae, Kim, Hyunwoo J.","roleAtSource":"Article author"},"source":{"title":"AgentGrad: Intervention-guided Prompt Optimization for Multi-Agent Systems","publisher":"arXiv cs.AI","format":"Article","url":"https://arxiv.org/abs/2609.08572","publishedAt":"2026-09-08","dateBasis":"Date basis not provided."},"evidence":[{"text":"The authors identify prompt selection without verifying its effect and lack of agent-level supervision as limitations in gradient extraction.","excerpt":"previous works select a prompt to update without verifying whether modifying it resolves the failure, and derive gradients without agent-level supervision","url":"https://arxiv.org/abs/2609.08572","relationship":"Source-stated"},{"text":"AgentGrad uses sequential intervention to identify which agent’s modification resolves a failure, then uses that agent’s modified output as supervision.","excerpt":"sequential intervention injects a hint into one agent at a time to identify the target agent whose modification resolves the failure","url":"https://arxiv.org/abs/2609.08572","relationship":"Source-stated"},{"text":"AgentGrad clusters semantically similar gradients and abstracts each cluster into a generalized gradient intended to capture shared corrective patterns.","excerpt":"clusters semantically similar gradients to prevent mixing unrelated failure modes, and abstracts each cluster into a generalized gradient","url":"https://arxiv.org/abs/2609.08572","relationship":"Source-stated"},{"text":"The authors report state-of-the-art performance on five MAS benchmarks and average reductions in optimization time and cost compared with the next-best baselines.","excerpt":"achieves state-of-the-art performance across five MAS benchmarks while reducing wall-clock optimization time by 2.5 times and optimization cost by 21.8% on average","url":"https://arxiv.org/abs/2609.08572","relationship":"Source-stated"}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2609.08572","checkedAt":"2026-11-10T04:50:09.070Z","items":[],"note":"No suitable media was found by the supported page extraction."},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"How do AgentGrad’s results vary across the five MAS benchmarks, and what evidence shows its prompts generalize beyond them?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Jaewon Chu and coauthors propose AgentGrad for multi-agent systems","summary":"Jaewon Chu and coauthors propose AgentGrad, a framework for optimizing prompts in multi-agent systems (MAS), where specialized agents work together and performance depends on each agent’s prompt. They argue that existing textual gradient methods may update a prompt without checking whether the change fixes a failure, and may combine gradients from unrelated failures. AgentGrad uses sequential intervention: it adds a hint to one agent at a time, then uses the modified output of an agent whose change resolves the failure to supervise gradient extraction. Semantic textual gradient abstraction groups similar gradients into generalized gradients that capture shared corrective patterns. The authors report state-of-the-art performance across five MAS benchmarks, with 2.5 times less wall-clock optimization time and 21.8% lower optimization cost on average than the next-best baselines. The approach ties prompt updates to the agent associated with a failure and groups similar corrections; whether that improves performance beyond the reported benchmarks is not established by the abstract. The paper was accepted to NeurIPS 2026.","question":"How do AgentGrad’s results vary across the five MAS benchmarks, and what evidence shows its prompts generalize beyond them?","claims":[{"text":"The authors identify prompt selection without verifying its effect and lack of agent-level supervision as limitations in gradient extraction.","excerpt":"previous works select a prompt to update without verifying whether modifying it resolves the failure, and derive gradients without agent-level supervision","support":"source-stated"},{"text":"AgentGrad uses sequential intervention to identify which agent’s modification resolves a failure, then uses that agent’s modified output as supervision.","excerpt":"sequential intervention injects a hint into one agent at a time to identify the target agent whose modification resolves the failure","support":"source-stated"},{"text":"AgentGrad clusters semantically similar gradients and abstracts each cluster into a generalized gradient intended to capture shared corrective patterns.","excerpt":"clusters semantically similar gradients to prevent mixing unrelated failure modes, and abstracts each cluster into a generalized gradient","support":"source-stated"},{"text":"The authors report state-of-the-art performance on five MAS benchmarks and average reductions in optimization time and cost compared with the next-best baselines.","excerpt":"achieves state-of-the-art performance across five MAS benchmarks while reducing wall-clock optimization time by 2.5 times and optimization cost by 21.8% on average","support":"source-stated"}],"limitations":["The evidence is the arXiv abstract; it does not provide benchmark-level results or details of the experimental measurements.","The reported performance and optimization savings are the authors’ claims about their experiments; the abstract does not establish real-world reliability or performance beyond the five MAS benchmarks."],"mediaId":"e121ff84f01d21f9fde526eb56f160dba769b549384e3795973328a4d0c6a81a","correctionReason":null,"id":"14f66bbfc639af860d443396c291d8906dae763f3ebd672785d390976d5f3986","publicationId":"e31cf00b-c446-425c-8979-e72749dcadaf","version":1,"sourceUrl":"https://arxiv.org/abs/2609.08572","sourceTitle":"AgentGrad: Intervention-guided Prompt Optimization for Multi-Agent Systems","publisher":"arXiv cs.AI","publishedAt":null,"originalPublished":{"value":"2026-09-08","precision":"day","basis":"publisher_citation_date"},"firstObservedAt":"2026-11-09T05:03:39.821Z","createdAt":"2026-11-10T03:46:31.867Z","historical":true,"entityIds":["arxiv"],"topicIds":["intelligence-and-discovery"],"media":[{"id":"e121ff84f01d21f9fde526eb56f160dba769b549384e3795973328a4d0c6a81a","kind":"article","url":"https://arxiv.org/abs/2609.08572","sourceUrl":"https://arxiv.org/abs/2609.08572","label":"Read original","provenance":"source_link","verification":{"status":"verified","checkedAt":"2026-11-10T03:16:37.161Z","method":"acquired_source","contentType":"text/html"},"recommendationReason":"Recommended by Jev for understanding this source; checked by Luna."}],"sourceMedia":{"status":"not_found","sourceUrl":"https://arxiv.org/abs/2609.08572","checkedAt":"2026-11-10T04:50:09.070Z","items":[],"note":"No suitable media was found by the supported page extraction."},"editorialDecision":{"outcome":"accept","reasonCodes":["material_development"],"rubricVersion":"stream-admission-v5","mediaId":"e121ff84f01d21f9fde526eb56f160dba769b549384e3795973328a4d0c6a81a","confidence":0.87,"provider":"typesafe","model":"jev-1.13.0","inputHash":"81be96dd3dadd128baf924359ebcd7d9bd590c967f372b38cd69306d60bed647","decidedAt":"2026-11-10T03:16:37.964Z"},"provenance":{"textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observationId":"48cb643edddb29d63eab6f4c60250d8b61e329b366102b57ff127e77f4f34441","retrievedAt":"2026-11-10T03:16:37.039Z","kind":"publisher-page","preparer":{"model":"gpt-6-luna","runId":"b8cd52b7-0ff5-4ae7-80fa-571ebeed65a7"},"reviewer":{"model":"gpt-6-luna","runId":"b7a7c37f-0a94-41f3-8d21-a27022a1f877"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-10T03:46:31.867Z","model":"anthropic/claude-opus-5.5"},"sourceAttribution":{"authors":[{"name":"Chu, Jaewon","url":null,"kind":"unknown"},{"name":"Seo, Jinwoo","url":null,"kind":"unknown"},{"name":"Cho, Jaewon","url":null,"kind":"unknown"},{"name":"Na, Jeehye","url":null,"kind":"unknown"},{"name":"Xiong, Yunyang","url":null,"kind":"unknown"},{"name":"Kim, Youngdae","url":null,"kind":"unknown"},{"name":"Kim, Hyunwoo J.","url":null,"kind":"unknown"}],"authorRole":"article-author","hosts":[],"status":"stated","authorRoster":{"truncated":false,"omitted":0,"assertionGroupsOmitted":0},"participation":[{"actor":{"name":"Chu, Jaewon","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Chu, Jaewon","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2609.08572","locator":"head/meta[name=author|citation_author]"},"assertionId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36:participation:0","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T04:50:09.070Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36","sourceUrl":"https://arxiv.org/abs/2609.08572","contentSha256":"8fba1ec8f95099e9041e160de6dd94a45ea8e4a619901564934097be61fafee5","textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observedAt":"2026-11-10T04:50:09.070Z"}},{"actor":{"name":"Seo, Jinwoo","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Seo, Jinwoo","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2609.08572","locator":"head/meta[name=author|citation_author]"},"assertionId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36:participation:1","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T04:50:09.070Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36","sourceUrl":"https://arxiv.org/abs/2609.08572","contentSha256":"8fba1ec8f95099e9041e160de6dd94a45ea8e4a619901564934097be61fafee5","textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observedAt":"2026-11-10T04:50:09.070Z"}},{"actor":{"name":"Cho, Jaewon","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Cho, Jaewon","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2609.08572","locator":"head/meta[name=author|citation_author]"},"assertionId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36:participation:2","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T04:50:09.070Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36","sourceUrl":"https://arxiv.org/abs/2609.08572","contentSha256":"8fba1ec8f95099e9041e160de6dd94a45ea8e4a619901564934097be61fafee5","textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observedAt":"2026-11-10T04:50:09.070Z"}},{"actor":{"name":"Na, Jeehye","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Na, Jeehye","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2609.08572","locator":"head/meta[name=author|citation_author]"},"assertionId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36:participation:3","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T04:50:09.070Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36","sourceUrl":"https://arxiv.org/abs/2609.08572","contentSha256":"8fba1ec8f95099e9041e160de6dd94a45ea8e4a619901564934097be61fafee5","textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observedAt":"2026-11-10T04:50:09.070Z"}},{"actor":{"name":"Xiong, Yunyang","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Xiong, Yunyang","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2609.08572","locator":"head/meta[name=author|citation_author]"},"assertionId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36:participation:4","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T04:50:09.070Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36","sourceUrl":"https://arxiv.org/abs/2609.08572","contentSha256":"8fba1ec8f95099e9041e160de6dd94a45ea8e4a619901564934097be61fafee5","textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observedAt":"2026-11-10T04:50:09.070Z"}},{"actor":{"name":"Kim, Youngdae","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Kim, Youngdae","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2609.08572","locator":"head/meta[name=author|citation_author]"},"assertionId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36:participation:5","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T04:50:09.070Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36","sourceUrl":"https://arxiv.org/abs/2609.08572","contentSha256":"8fba1ec8f95099e9041e160de6dd94a45ea8e4a619901564934097be61fafee5","textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observedAt":"2026-11-10T04:50:09.070Z"}},{"actor":{"name":"Kim, Hyunwoo J.","url":null,"kind":"unknown","actorId":null},"role":"author","scope":{"kind":"work-credit","locator":null},"literalCredit":"Kim, Hyunwoo J.","status":"stated","provenance":{"basis":"head-meta","sourceUrl":"https://arxiv.org/abs/2609.08572","locator":"head/meta[name=author|citation_author]"},"assertionId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36:participation:6","identity":{"version":"entity-identity-v1","status":"unresolved","actorId":null,"canonicalActorId":null,"candidateActorIds":[],"assertions":[],"sourceObservedAt":"2026-11-10T04:50:09.070Z","knowledgeAsOf":null},"sourceVersion":{"snapshotId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36","sourceUrl":"https://arxiv.org/abs/2609.08572","contentSha256":"8fba1ec8f95099e9041e160de6dd94a45ea8e4a619901564934097be61fafee5","textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observedAt":"2026-11-10T04:50:09.070Z"}}],"participationCoverage":{"total":7,"returned":7,"complete":true,"fullAssertions":"source-metadata-endpoint"},"distribution":null,"publisher":null,"sourceUrl":"https://arxiv.org/abs/2609.08572","snapshotId":"6dfe969e3cb3fd51f60b74bdade91bdd33a43967577857d58ea799886c6fcb36","contentSha256":"8fba1ec8f95099e9041e160de6dd94a45ea8e4a619901564934097be61fafee5","textHash":"373adf2d672ad541d8d91011e965c8129f919c815615d977c6dfb02c9e57369e","observedAt":"2026-11-10T04:50:09.070Z","basis":"publisher-page-metadata","sourceLinks":49,"linksTruncated":false,"metadataUrl":"/api/pipeline/stream/14f66bbfc639af860d443396c291d8906dae763f3ebd672785d390976d5f3986/source"}}}],"audits":[{"publicationId":"e31cf00b-c446-425c-8979-e72749dcadaf","verdict":"uncertain","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading correctly attributes AgentGrad and its claims to Jaewon Chu and coauthors, matching the arXiv author list. Results are presented as the authors' reports, not as established facts.","excerpt":"Authors:Jaewon Chu, Jinwoo Seo, Jaewon Cho, Jeehye Na, Yunyang Xiong, Youngdae Kim, Hyunwoo J. Kim"},{"dimension":"summary","status":"uncertain","explanation":"The summary is faithful, but a central term is not explained. It uses 'textual gradient methods' and 'gradients' without saying what they are. The source defines them in plain language as methods that guide prompt updates using natural-language feedback. Without that, a non-specialist cannot follow the critique or AgentGrad's fix. Smallest repair: add a short gloss, for example that textual gradients are natural-language feedback used to revise prompts, or confirm a term-library tooltip exists. The title, question and other content are accurate.","excerpt":"textual gradient methods that guide prompt updates using natural-language feedback have emerged as a leading paradigm"},{"dimension":"date","status":"pass","explanation":"The originalPublished value, 2026-09-08, matches the v1 submission date. It is not the date of the later v3 revision (6 Nov 2026) or the feed date.","excerpt":"Submitted on 8 Sep 2026 (v1), last revised 6 Nov 2026 (this version, v3)"},{"dimension":"support","status":"pass","explanation":"Each claim matches the abstract, including the method steps, the five-benchmark result, the 2.5x time reduction, the 21.8% average cost reduction and the NeurIPS acceptance. The interpretive sentence is properly hedged and does not overstate generalization. The question follows from the evidence without assuming an answer.","excerpt":"achieves state-of-the-art performance across five MAS benchmarks while reducing wall-clock optimization time by 2.5 times and optimization cost by 21.8% on average"},{"dimension":"media","status":"pass","explanation":"The selected media is the verified direct link to the arXiv abstract page, which is the source itself.","excerpt":null}],"createdAt":"2026-11-10T03:46:31.867Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}