[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"blog":3},{"title":4,"desc":5,"bannerImg":6,"date":7,"links":8,"description":5,"content":9,"tag1":729,"tag2":730,"resLinks":733},"Introducing Chain-of-Agents: A New Paradigm for Agent Foundation Model","Discover Chain-of-Agents (CoA), a breakthrough framework for training powerful and efficient Agent Foundation Models. Learn how AFMs achieve state-of-the-art results on 20+ benchmarks while cutting costs. Explore the open-source models and code.","https:\u002F\u002Fdoxhub.s3.us-east-1.amazonaws.com\u002F2077ai\u002FBanner_blog\u002Fbanner_chain%20of%20agents.png","2025-09-03","{\"github\":\"https:\u002F\u002Fgithub.com\u002FOPPO-PersonalAI\u002FAgent_Foundation_Models\",\"huggingface\":\"https:\u002F\u002Fhuggingface.co\u002Fcollections\u002FPersonalAILab\u002Fafm-datasets-6892140eaad360ea5ccdcde1\", \"arxiv\":\"https:\u002F\u002Farxiv.org\u002Fabs\u002F2508.13167\",\"homepage\":\"https:\u002F\u002Fchain-of-agents-afm.github.io\u002F\"}",{"data":10,"body":13,"toc":716},{"title":11,"description":12},"Redefining AI Collaboration: Introducing Chain-of-Agents and Agent Foundation Models","The field of artificial intelligence is moving at an incredible pace, with AI agents capable of tackling complex, multi-step tasks becoming a reality. At 2077AI, we are thrilled to have contributed to a significant leap forward in this domain. We're excited to spotlight the groundbreaking latest paper and open-source release from our collaborators at OPPO's Personalized AI Lab and other leading researchers: Chain-of-Agents (CoA), a new paradigm that trains Agent Foundation Models (AFM) for unparalleled performance and efficiency.",{"type":14,"children":15},"root",[16,32,75,85,150,160,173,197,207,345,360,370,381,391,403,416,426,437,447,458,468,479,489,501,511,652,676,687,697,707],{"type":17,"tag":18,"props":19,"children":24},"element","h1",{"className":20,"lexical-key":22,"id":23},[21],"heading__h1","387","redefining-ai-collaboration-introducing-chain-of-agents-and-agent-foundation-models",[25],{"type":17,"tag":26,"props":27,"children":29},"span",{"style":28},"white-space: pre-wrap;",[30],{"type":31,"value":11},"text",{"type":17,"tag":33,"props":34,"children":38},"p",{"className":35,"lexical-key":37},[36],"doxhub-editor-paragraph","389",[39,44,56,61,70],{"type":17,"tag":26,"props":40,"children":41},{"style":28},[42],{"type":31,"value":43},"The field of artificial intelligence is moving at an incredible pace, with AI agents capable of tackling complex, multi-step tasks becoming a reality. At 2077AI, we are thrilled to have contributed to a significant leap forward in this domain. We're excited to spotlight the groundbreaking latest paper and open-source release from our collaborators at OPPO's Personalized AI Lab and other leading researchers: ",{"type":17,"tag":45,"props":46,"children":47},"b",{},[48],{"type":17,"tag":49,"props":50,"children":53},"strong",{"className":51,"style":28},[52],"text__bold",[54],{"type":31,"value":55},"Chain-of-Agents (CoA)",{"type":17,"tag":26,"props":57,"children":58},{"style":28},[59],{"type":31,"value":60},", a new paradigm that trains ",{"type":17,"tag":45,"props":62,"children":63},{},[64],{"type":17,"tag":49,"props":65,"children":67},{"className":66,"style":28},[52],[68],{"type":31,"value":69},"Agent Foundation Models (AFM)",{"type":17,"tag":26,"props":71,"children":72},{"style":28},[73],{"type":31,"value":74}," for unparalleled performance and efficiency.",{"type":17,"tag":33,"props":76,"children":79},{"className":77,"lexical-key":78},[36],"395",[80],{"type":17,"tag":26,"props":81,"children":82},{"style":28},[83],{"type":31,"value":84},"While multi-agent systems (MAS) have shown promise, they often face critical limitations:",{"type":17,"tag":86,"props":87,"children":90},"ul",{"className":88},[89],"doxhub-editor-ul",[91,112,131],{"type":17,"tag":92,"props":93,"children":97},"li",{"value":94,"className":95},"1",[96],"doxhub-editor-list-item",[98,107],{"type":17,"tag":45,"props":99,"children":100},{},[101],{"type":17,"tag":49,"props":102,"children":104},{"className":103,"style":28},[52],[105],{"type":31,"value":106},"High Computational Costs:",{"type":17,"tag":26,"props":108,"children":109},{"style":28},[110],{"type":31,"value":111}," Frequent and redundant communication between agents makes them slow and expensive.",{"type":17,"tag":92,"props":113,"children":116},{"value":114,"className":115},"2",[96],[117,126],{"type":17,"tag":45,"props":118,"children":119},{},[120],{"type":17,"tag":49,"props":121,"children":123},{"className":122,"style":28},[52],[124],{"type":31,"value":125},"Limited Generalization:",{"type":17,"tag":26,"props":127,"children":128},{"style":28},[129],{"type":31,"value":130}," Adapting to new tasks requires extensive and costly manual prompt engineering and workflow design.",{"type":17,"tag":92,"props":132,"children":135},{"value":133,"className":134},"3",[96],[136,145],{"type":17,"tag":45,"props":137,"children":138},{},[139],{"type":17,"tag":49,"props":140,"children":142},{"className":141,"style":28},[52],[143],{"type":31,"value":144},"Lack of Learnability:",{"type":17,"tag":26,"props":146,"children":147},{"style":28},[148],{"type":31,"value":149}," Most systems can't learn and improve from data, hitting a performance ceiling defined by their initial design.",{"type":17,"tag":33,"props":151,"children":154},{"className":152,"lexical-key":153},[36],"407",[155],{"type":17,"tag":26,"props":156,"children":157},{"style":28},[158],{"type":31,"value":159},"The Chain-of-Agents paradigm was created to solve these challenges head-on.",{"type":17,"tag":161,"props":162,"children":167},"h2",{"className":163,"lexical-key":165,"id":166},[164],"heading__h2","409","chain-of-agents-a-new-paradigm-for-native-collaboration",[168],{"type":17,"tag":26,"props":169,"children":170},{"style":28},[171],{"type":31,"value":172},"Chain-of-Agents: A New Paradigm for Native Collaboration",{"type":17,"tag":33,"props":174,"children":177},{"className":175,"lexical-key":176},[36],"411",[178,183,192],{"type":17,"tag":26,"props":179,"children":180},{"style":28},[181],{"type":31,"value":182},"Instead of relying on multiple, separate models governed by complex external frameworks, Chain-of-Agents (CoA) enables a ",{"type":17,"tag":45,"props":184,"children":185},{},[186],{"type":17,"tag":49,"props":187,"children":189},{"className":188,"style":28},[52],[190],{"type":31,"value":191},"single, end-to-end model",{"type":17,"tag":26,"props":193,"children":194},{"style":28},[195],{"type":31,"value":196}," to simulate multi-agent collaboration internally.",{"type":17,"tag":33,"props":198,"children":201},{"className":199,"lexical-key":200},[36],"415",[202],{"type":17,"tag":26,"props":203,"children":204},{"style":28},[205],{"type":31,"value":206},"CoA employs a hierarchical agent architecture that can be dynamically activated within the model:",{"type":17,"tag":86,"props":208,"children":210},{"className":209},[89],[211,285],{"type":17,"tag":92,"props":212,"children":214},{"value":94,"className":213},[96],[215,224,229,238,243,252,257,266,271,280],{"type":17,"tag":45,"props":216,"children":217},{},[218],{"type":17,"tag":49,"props":219,"children":221},{"className":220,"style":28},[52],[222],{"type":31,"value":223},"Role-playing Agents:",{"type":17,"tag":26,"props":225,"children":226},{"style":28},[227],{"type":31,"value":228}," These agents handle the reasoning process. They include a ",{"type":17,"tag":45,"props":230,"children":231},{},[232],{"type":17,"tag":49,"props":233,"children":235},{"className":234,"style":28},[52],[236],{"type":31,"value":237},"Thinking Agent",{"type":17,"tag":26,"props":239,"children":240},{"style":28},[241],{"type":31,"value":242}," for analysis, a ",{"type":17,"tag":45,"props":244,"children":245},{},[246],{"type":17,"tag":49,"props":247,"children":249},{"className":248,"style":28},[52],[250],{"type":31,"value":251},"Plan Agent",{"type":17,"tag":26,"props":253,"children":254},{"style":28},[255],{"type":31,"value":256}," for strategy, a ",{"type":17,"tag":45,"props":258,"children":259},{},[260],{"type":17,"tag":49,"props":261,"children":263},{"className":262,"style":28},[52],[264],{"type":31,"value":265},"Reflection Agent",{"type":17,"tag":26,"props":267,"children":268},{"style":28},[269],{"type":31,"value":270}," for self-correction, and a ",{"type":17,"tag":45,"props":272,"children":273},{},[274],{"type":17,"tag":49,"props":275,"children":277},{"className":276,"style":28},[52],[278],{"type":31,"value":279},"Verification Agent",{"type":17,"tag":26,"props":281,"children":282},{"style":28},[283],{"type":31,"value":284}," for confirming results.",{"type":17,"tag":92,"props":286,"children":288},{"value":114,"className":287},[96],[289,298,303,312,317,326,331,340],{"type":17,"tag":45,"props":290,"children":291},{},[292],{"type":17,"tag":49,"props":293,"children":295},{"className":294,"style":28},[52],[296],{"type":31,"value":297},"Tool Agents:",{"type":17,"tag":26,"props":299,"children":300},{"style":28},[301],{"type":31,"value":302}," These agents execute specific actions, such as a ",{"type":17,"tag":45,"props":304,"children":305},{},[306],{"type":17,"tag":49,"props":307,"children":309},{"className":308,"style":28},[52],[310],{"type":31,"value":311},"Search Agent",{"type":17,"tag":26,"props":313,"children":314},{"style":28},[315],{"type":31,"value":316}," for finding information, a ",{"type":17,"tag":45,"props":318,"children":319},{},[320],{"type":17,"tag":49,"props":321,"children":323},{"className":322,"style":28},[52],[324],{"type":31,"value":325},"Crawl Agent",{"type":17,"tag":26,"props":327,"children":328},{"style":28},[329],{"type":31,"value":330}," for accessing web content, and a ",{"type":17,"tag":45,"props":332,"children":333},{},[334],{"type":17,"tag":49,"props":335,"children":337},{"className":336,"style":28},[52],[338],{"type":31,"value":339},"Code Agent",{"type":17,"tag":26,"props":341,"children":342},{"style":28},[343],{"type":31,"value":344}," for writing and executing code.",{"type":17,"tag":346,"props":347,"children":348},"figure",{},[349,355],{"type":17,"tag":350,"props":351,"children":354},"img",{"src":352,"alt":353},"https:\u002F\u002Fdoxhub.s3.us-east-1.amazonaws.com\u002F2077ai\u002F20250903\u002Fchainofagent-01.png","\n    The core technical pillars of the AFM framework\n  ",[],{"type":17,"tag":356,"props":357,"children":358},"figcaption",{},[359],{"type":31,"value":353},{"type":17,"tag":33,"props":361,"children":364},{"className":362,"lexical-key":363},[36],"440",[365],{"type":17,"tag":26,"props":366,"children":367},{"style":28},[368],{"type":31,"value":369},"By integrating these roles into one model, AFM eliminates the need for complex prompt engineering and drastically reduces communication overhead, leading to a more efficient and powerful system.",{"type":17,"tag":161,"props":371,"children":375},{"className":372,"lexical-key":373,"id":374},[164],"442","how-we-built-afm-a-three-stage-training-framework",[376],{"type":17,"tag":26,"props":377,"children":378},{"style":28},[379],{"type":31,"value":380},"How We Built AFM: A Three-Stage Training Framework",{"type":17,"tag":33,"props":382,"children":385},{"className":383,"lexical-key":384},[36],"444",[386],{"type":17,"tag":26,"props":387,"children":388},{"style":28},[389],{"type":31,"value":390},"Creating a model with these native agentic capabilities required a novel training framework that combines multi-agent distillation with reinforcement learning.",{"type":17,"tag":346,"props":392,"children":393},{},[394,399],{"type":17,"tag":350,"props":395,"children":398},{"src":396,"alt":397},"https:\u002F\u002Fdoxhub.s3.us-east-1.amazonaws.com\u002F2077ai\u002F20250903\u002Fchainofagent-02.png","\n    The AFM training framework\n  ",[],{"type":17,"tag":356,"props":400,"children":401},{},[402],{"type":31,"value":397},{"type":17,"tag":404,"props":405,"children":410},"h3",{"className":406,"lexical-key":408,"id":409},[407],"heading__h3","448","_1-trajectory-acquisition-distillation",[411],{"type":17,"tag":26,"props":412,"children":413},{"style":28},[414],{"type":31,"value":415},"1. Trajectory Acquisition & Distillation",{"type":17,"tag":33,"props":417,"children":420},{"className":418,"lexical-key":419},[36],"450",[421],{"type":17,"tag":26,"props":422,"children":423},{"style":28},[424],{"type":31,"value":425},"The process starts by collecting a diverse set of tasks across web navigation, math, and coding. An advanced multi-agent system, OAgents, is used to solve these tasks. The successful solution paths, or \"trajectories,\" are then distilled and converted into a format compatible with the CoA paradigm. This creates a high-quality dataset for the next stage.",{"type":17,"tag":404,"props":427,"children":431},{"className":428,"lexical-key":429,"id":430},[407],"452","_2-supervised-fine-tuning-sft",[432],{"type":17,"tag":26,"props":433,"children":434},{"style":28},[435],{"type":31,"value":436},"2. Supervised Fine-Tuning (SFT)",{"type":17,"tag":33,"props":438,"children":441},{"className":439,"lexical-key":440},[36],"454",[442],{"type":17,"tag":26,"props":443,"children":444},{"style":28},[445],{"type":31,"value":446},"The base Large Language Model (LLM) is fine-tuned on these distilled trajectories. This SFT stage teaches the model the foundational patterns of Chain-of-Agents reasoning, embedding the ability to plan, act, and reflect.",{"type":17,"tag":404,"props":448,"children":452},{"className":449,"lexical-key":450,"id":451},[407],"456","_3-agent-reinforcement-learning-rl",[453],{"type":17,"tag":26,"props":454,"children":455},{"style":28},[456],{"type":31,"value":457},"3. Agent Reinforcement Learning (RL)",{"type":17,"tag":33,"props":459,"children":462},{"className":460,"lexical-key":461},[36],"458",[463],{"type":17,"tag":26,"props":464,"children":465},{"style":28},[466],{"type":31,"value":467},"To elevate the model from simply mimicking trajectories to developing an optimal strategy, we employ reinforcement learning. The model performs tasks, and its actions are evaluated by a sophisticated reward model. This reward system uses rule-based checks for verifiable tasks (like passing test cases in code) and an \"LLM-as-a-Judge\" to assess reasoning coherence, tool efficiency, and answer precision. The policy is continually updated, sharpening the model’s problem-solving strategies, especially for the most challenging tasks.",{"type":17,"tag":161,"props":469,"children":473},{"className":470,"lexical-key":471,"id":472},[164],"460","unprecedented-performance-across-the-board",[474],{"type":17,"tag":26,"props":475,"children":476},{"style":28},[477],{"type":31,"value":478},"Unprecedented Performance Across the Board",{"type":17,"tag":33,"props":480,"children":483},{"className":481,"lexical-key":482},[36],"462",[484],{"type":17,"tag":26,"props":485,"children":486},{"style":28},[487],{"type":31,"value":488},"The results speak for themselves. Agent Foundation Models have set a new state-of-the-art across nearly 20 complex agent benchmarks.",{"type":17,"tag":346,"props":490,"children":491},{},[492,497],{"type":17,"tag":350,"props":493,"children":496},{"src":494,"alt":495},"https:\u002F\u002Fdoxhub.s3.us-east-1.amazonaws.com\u002F2077ai\u002F20250903\u002Fchainofagent-03.png","\n    AFM establishes its leadership across four challenging agent benchmarks\n",[],{"type":17,"tag":356,"props":498,"children":499},{},[500],{"type":31,"value":495},{"type":17,"tag":33,"props":502,"children":505},{"className":503,"lexical-key":504},[36],"466",[506],{"type":17,"tag":26,"props":507,"children":508},{"style":28},[509],{"type":31,"value":510},"As shown in the benchmarks:",{"type":17,"tag":86,"props":512,"children":514},{"className":513},[89],[515,552,616],{"type":17,"tag":92,"props":516,"children":518},{"value":94,"className":517},[96],[519,524,533,538,547],{"type":17,"tag":26,"props":520,"children":521},{"style":28},[522],{"type":31,"value":523},"On ",{"type":17,"tag":45,"props":525,"children":526},{},[527],{"type":17,"tag":49,"props":528,"children":530},{"className":529,"style":28},[52],[531],{"type":31,"value":532},"GAIA",{"type":17,"tag":26,"props":534,"children":535},{"style":28},[536],{"type":31,"value":537},", a general AI assistant benchmark, AFM achieves a score of ",{"type":17,"tag":45,"props":539,"children":540},{},[541],{"type":17,"tag":49,"props":542,"children":544},{"className":543,"style":28},[52],[545],{"type":31,"value":546},"55.3",{"type":17,"tag":26,"props":548,"children":549},{"style":28},[550],{"type":31,"value":551},", surpassing previous leading models.",{"type":17,"tag":92,"props":553,"children":555},{"value":114,"className":554},[96],[556,561,570,575,584,589,598,602,611],{"type":17,"tag":26,"props":557,"children":558},{"style":28},[559],{"type":31,"value":560},"For complex web navigation, AFM scores ",{"type":17,"tag":45,"props":562,"children":563},{},[564],{"type":17,"tag":49,"props":565,"children":567},{"className":566,"style":28},[52],[568],{"type":31,"value":569},"11.1",{"type":17,"tag":26,"props":571,"children":572},{"style":28},[573],{"type":31,"value":574}," on ",{"type":17,"tag":45,"props":576,"children":577},{},[578],{"type":17,"tag":49,"props":579,"children":581},{"className":580,"style":28},[52],[582],{"type":31,"value":583},"BrowseComp",{"type":17,"tag":26,"props":585,"children":586},{"style":28},[587],{"type":31,"value":588}," and ",{"type":17,"tag":45,"props":590,"children":591},{},[592],{"type":17,"tag":49,"props":593,"children":595},{"className":594,"style":28},[52],[596],{"type":31,"value":597},"18.0",{"type":17,"tag":26,"props":599,"children":600},{"style":28},[601],{"type":31,"value":574},{"type":17,"tag":45,"props":603,"children":604},{},[605],{"type":17,"tag":49,"props":606,"children":608},{"className":607,"style":28},[52],[609],{"type":31,"value":610},"HLE",{"type":17,"tag":26,"props":612,"children":613},{"style":28},[614],{"type":31,"value":615},", demonstrating superior tool usage and planning.",{"type":17,"tag":92,"props":617,"children":619},{"value":133,"className":618},[96],[620,625,634,638,647],{"type":17,"tag":26,"props":621,"children":622},{"style":28},[623],{"type":31,"value":624},"In the domain of advanced mathematical reasoning, AFM reaches a commanding ",{"type":17,"tag":45,"props":626,"children":627},{},[628],{"type":17,"tag":49,"props":629,"children":631},{"className":630,"style":28},[52],[632],{"type":31,"value":633},"59.8",{"type":17,"tag":26,"props":635,"children":636},{"style":28},[637],{"type":31,"value":574},{"type":17,"tag":45,"props":639,"children":640},{},[641],{"type":17,"tag":49,"props":642,"children":644},{"className":643,"style":28},[52],[645],{"type":31,"value":646},"AIME25",{"type":17,"tag":26,"props":648,"children":649},{"style":28},[650],{"type":31,"value":651},".",{"type":17,"tag":33,"props":653,"children":656},{"className":654,"lexical-key":655},[36],"491",[657,662,671],{"type":17,"tag":26,"props":658,"children":659},{"style":28},[660],{"type":31,"value":661},"Beyond raw performance, AFM is also remarkably efficient. In our analysis, it reduced the number of tokens required for inference by up to ",{"type":17,"tag":45,"props":663,"children":664},{},[665],{"type":17,"tag":49,"props":666,"children":668},{"className":667,"style":28},[52],[669],{"type":31,"value":670},"85.5%",{"type":17,"tag":26,"props":672,"children":673},{"style":28},[674],{"type":31,"value":675}," compared to traditional multi-agent frameworks, delivering top-tier results at a fraction of the computational cost.",{"type":17,"tag":161,"props":677,"children":681},{"className":678,"lexical-key":679,"id":680},[164],"495","the-future-is-agentic-and-its-here",[682],{"type":17,"tag":26,"props":683,"children":684},{"style":28},[685],{"type":31,"value":686},"The Future is Agentic, and It's Here",{"type":17,"tag":33,"props":688,"children":691},{"className":689,"lexical-key":690},[36],"497",[692],{"type":17,"tag":26,"props":693,"children":694},{"style":28},[695],{"type":31,"value":696},"We are standing at the edge of a new frontier in artificial intelligence. The introduction of the Chain-of-Agents paradigm and Agent Foundation Models is more than just an academic breakthrough; it represents a fundamental shift in how we build intelligent systems. We are moving beyond the era of rigid, manually-coded agents and stepping into a future of dynamic, learning entities that can reason, adapt, and solve problems with unprecedented autonomy.",{"type":17,"tag":33,"props":698,"children":701},{"className":699,"lexical-key":700},[36],"499",[702],{"type":17,"tag":26,"props":703,"children":704},{"style":28},[705],{"type":31,"value":706},"At 2077AI, we are incredibly proud to have played a part in this visionary project. We believe this work doesn't just lay the groundwork — it unlocks the door to a future where AI agents become true partners in discovery, creation, and human progress. By making this entire framework fully open-source, we are extending an invitation to every developer, researcher, and dreamer to join us on this journey. Let's build the future of AI, together.",{"type":17,"tag":33,"props":708,"children":711},{"className":709,"lexical-key":710},[36],"501",[712],{"type":17,"tag":713,"props":714,"children":715},"br",{},[],{"title":717,"searchDepth":718,"depth":718,"links":719},"",2,[720,721,727,728],{"id":166,"depth":718,"text":172},{"id":374,"depth":718,"text":380,"children":722},[723,725,726],{"id":409,"depth":724,"text":415},3,{"id":430,"depth":724,"text":436},{"id":451,"depth":724,"text":457},{"id":472,"depth":718,"text":478},{"id":680,"depth":718,"text":686},"model",[731,732],"agent","llm",{"homepage":734,"arxiv":735,"github":736,"huggingface":737},"https:\u002F\u002Fchain-of-agents-afm.github.io\u002F","https:\u002F\u002Farxiv.org\u002Fabs\u002F2508.13167","https:\u002F\u002Fgithub.com\u002FOPPO-PersonalAI\u002FAgent_Foundation_Models","https:\u002F\u002Fhuggingface.co\u002Fcollections\u002FPersonalAILab\u002Fafm-datasets"]