AgentGrad addresses the key limitations of existing textual gradient methods in both the extraction and aggregation stages for multi-agent prompt optimization. By introducing sequential intervention and semantic textual gradient abstraction, it substantially improves both optimization performance and speed.</p>\n","updatedAt":"2026-09-10T11:13:17.885Z","author":{"_id":"671b68eaf50df4c3c72ea53b","avatarUrl":"/avatars/c33d8baffab4e4d715c5a6aa3a85b55b.svg","fullname":"Jaewonchu","name":"allonsy07","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8496937155723572},"editors":["allonsy07"],"editorAvatarUrls":["/avatars/c33d8baffab4e4d715c5a6aa3a85b55b.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2609.08572","authors":[{"_id":"6aa18d8ad8c54e38c0a36aac","user":{"_id":"671b68eaf50df4c3c72ea53b","avatarUrl":"/avatars/c33d8baffab4e4d715c5a6aa3a85b55b.svg","isPro":false,"fullname":"Jaewonchu","user":"allonsy07","type":"user","name":"allonsy07"},"name":"Jaewon Chu","status":"claimed_verified","statusLastChangedAt":"2026-09-10T08:45:05.177Z","hidden":false},{"_id":"6aa18d8ad8c54e38c0a36aad","name":"Jinwoo Seo","hidden":false},{"_id":"6aa18d8ad8c54e38c0a36aae","name":"Jaewon Cho","hidden":false},{"_id":"6aa18d8ad8c54e38c0a36aaf","name":"Jeehye Na","hidden":false},{"_id":"6aa18d8ad8c54e38c0a36ab0","user":{"_id":"65304b62e7535baecd85d080","avatarUrl":"/avatars/6e546c7d1414bd92c5a7c8d8c404de92.svg","isPro":false,"fullname":"Yunyang Xiong","user":"yunyangx","type":"user","name":"yunyangx"},"name":"Yunyang Xiong","status":"claimed_verified","statusLastChangedAt":"2026-09-10T08:45:05.169Z","hidden":false},{"_id":"6aa18d8ad8c54e38c0a36ab1","name":"Youngdae Kim","hidden":false},{"_id":"6aa18d8ad8c54e38c0a36ab2","name":"Hyunwoo J. Kim","hidden":false}],"publishedAt":"2026-09-08T00:00:00.000Z","submittedOnDailyAt":"2026-09-10T00:00:00.000Z","title":"AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems","submittedOnDailyBy":{"_id":"671b68eaf50df4c3c72ea53b","avatarUrl":"/avatars/c33d8baffab4e4d715c5a6aa3a85b55b.svg","isPro":false,"fullname":"Jaewonchu","user":"allonsy07","type":"user","name":"allonsy07"},"summary":"Large language model (LLM)-based multi-agent systems (MAS) achieve strong performance by employing specialized multiple agents, yet their performance depends on the prompt design of each agent. For MAS prompt optimization, textual gradient methods that guide prompt updates using natural-language feedback have emerged as a leading paradigm. In this paper, we identify limitations in two stages of existing textual gradient approaches: gradient extraction and gradient aggregation. In gradient extraction, previous works select a target prompt without verifying whether modifying it resolves the failure, and derive gradients without agent-level supervision over the corresponding agent's intermediate output. In gradient aggregation, individual gradients are randomly grouped and concatenated, often mixing unrelated failure modes and producing prompts that fail to generalize. To address these limitations, we propose AgentGrad, a prompt optimization framework for multi-agent systems based on sequential intervention and semantic textual gradient abstraction. For each failure, sequential intervention modifies the behavior of one agent at a time to identify the target agent whose modification resolves the failure. The modified output of the target agent then serves as agent-level supervision for extracting a fine-grained gradient. Semantic textual gradient abstraction clusters semantically similar gradients to prevent mixing unrelated failure modes, and abstracts each cluster into a generalized gradient that captures the shared corrective pattern. Experimental results show that AgentGrad achieves state-of-the-art performance across five MAS benchmarks and reduces wall-clock optimization time by 2.5times on average compared to the next-fastest baseline.","upvotes":49,"discussionId":"6aa18d8ad8c54e38c0a36ab3","ai_summary":"AgentGrad improves multi-agent prompt optimization by identifying target agents through sequential intervention and clustering gradients semantically to avoid mixing unrelated errors.","ai_keywords":["multi-agent systems","textual gradient","gradient extraction","gradient aggregation","sequential intervention","semantic textual gradient abstraction","AgentGrad"],"ai_summary_model":"thinkingmachines/Inkling-Small"},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"65304b62e7535baecd85d080","avatarUrl":"/avatars/6e546c7d1414bd92c5a7c8d8c404de92.svg","isPro":false,"fullname":"Yunyang Xiong","user":"yunyangx","type":"user"},{"_id":"671b68eaf50df4c3c72ea53b","avatarUrl":"/avatars/c33d8baffab4e4d715c5a6aa3a85b55b.svg","isPro":false,"fullname":"Jaewonchu","user":"allonsy07","type":"user"},{"_id":"6687e814e977211be125fc7e","avatarUrl":"/avatars/38d8ad642aa7fd708206a75fd521ce27.svg","isPro":true,"fullname":"Jeehye Na","user":"sonicdog00","type":"user"},{"_id":"63d203f68eaa4831006af78a","avatarUrl":"/avatars/b6ef0b0bfaa6acdc1c6f967a0a5dec1f.svg","isPro":false,"fullname":"choijoonmyung","user":"joonmyung","type":"user"},{"_id":"651aa81811f562eb7f01efa3","avatarUrl":"/avatars/601222dd929cd4a770101e4ce731b5cb.svg","isPro":false,"fullname":"taehoon song","user":"turtle124","type":"user"},{"_id":"64b4c344ee7a5f18251d11ac","avatarUrl":"/avatars/8fabaa335d84388524519e8758978c6d.svg","isPro":false,"fullname":"Jinyoung Kim","user":"jinyoungkim","type":"user"},{"_id":"65fccc784d36be78e6834fb5","avatarUrl":"/avatars/a2024eeba2e2ee6a3e4e490d5d638090.svg","isPro":false,"fullname":"seoung choi","user":"choisw0823","type":"user"},{"_id":"65d089376236ca85a488d431","avatarUrl":"/avatars/98b691f3e2c6eee475e73d1ea7f47cea.svg","isPro":false,"fullname":"Jayla Cho","user":"chojaework","type":"user"},{"_id":"637db20b64686bd9011f51ba","avatarUrl":"/avatars/3b57ccd35d856958a84db6f0dd4e405e.svg","isPro":false,"fullname":"Taehoon Lee","user":"dlxogns183","type":"user"},{"_id":"69d39c2704bf200050b58d79","avatarUrl":"/avatars/5b2f6605c37565228b67f62f04aa82f2.svg","isPro":false,"fullname":"Kyujin Lee","user":"KyujinL","type":"user"},{"_id":"662b7fee3bdfe519489403d0","avatarUrl":"/avatars/5545ebff1cde05a3cbd76efe3ae4c07c.svg","isPro":true,"fullname":"minseok joo","user":"zoomkinseok","type":"user"},{"_id":"6078e2e7489fc71534e91be1","avatarUrl":"/avatars/1dbe35c9c92fc4e373de6647caa2e236.svg","isPro":false,"fullname":"Jihwan park","user":"Jhp","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":3,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2609/2609.08572.md","query":{}}">
AgentGrad: Intervention-guided Prompt Optimization for Multi Agent Systems
Abstract
AgentGrad improves multi-agent prompt optimization by identifying target agents through sequential intervention and clustering gradients semantically to avoid mixing unrelated errors.
Large language model (LLM)-based multi-agent systems (MAS) achieve strong performance by employing specialized multiple agents, yet their performance depends on the prompt design of each agent. For MAS prompt optimization, textual gradient methods that guide prompt updates using natural-language feedback have emerged as a leading paradigm. In this paper, we identify limitations in two stages of existing textual gradient approaches: gradient extraction and gradient aggregation. In gradient extraction, previous works select a target prompt without verifying whether modifying it resolves the failure, and derive gradients without agent-level supervision over the corresponding agent's intermediate output. In gradient aggregation, individual gradients are randomly grouped and concatenated, often mixing unrelated failure modes and producing prompts that fail to generalize. To address these limitations, we propose AgentGrad, a prompt optimization framework for multi-agent systems based on sequential intervention and semantic textual gradient abstraction. For each failure, sequential intervention modifies the behavior of one agent at a time to identify the target agent whose modification resolves the failure. The modified output of the target agent then serves as agent-level supervision for extracting a fine-grained gradient. Semantic textual gradient abstraction clusters semantically similar gradients to prevent mixing unrelated failure modes, and abstracts each cluster into a generalized gradient that captures the shared corrective pattern. Experimental results show that AgentGrad achieves state-of-the-art performance across five MAS benchmarks and reduces wall-clock optimization time by 2.5times on average compared to the next-fastest baseline.
Community
AgentGrad addresses the key limitations of existing textual gradient methods in both the extraction and aggregation stages for multi-agent prompt optimization. By introducing sequential intervention and semantic textual gradient abstraction, it substantially improves both optimization performance and speed.
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2609.08572 in a model README.md to link it from this page.
Cite arxiv.org/abs/2609.08572 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2609.08572 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.