[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-post-\u002Fnews\u002Fmeta-hyperagents-self-modification-metacognitive-autonomous-learning":3,"news-related-en-\u002Fnews\u002Fmeta-hyperagents-self-modification-metacognitive-autonomous-learning":451,"hreflang-meta-hyperagents-self-modification-metacognitive":471},{"doc":4,"redirect":8},{"id":5,"title":6,"author":7,"author_avatar":8,"body":9,"category":426,"date":427,"description":428,"extension":429,"image":430,"meta":431,"navigation":432,"path":433,"seo":434,"seoDescription":435,"seoTitle":436,"stem":437,"tags":438,"translationKey":449,"updatedAt":8,"__hash__":450},"news_en\u002Fnews\u002Fmeta-hyperagents-self-modification-metacognitive-autonomous-learning.md","Meta Releases HyperAgents: The AI That Improves Its Own Way of Improving — Metacognitive Self-Modification Explained","Paul Lefizelier",null,{"type":10,"value":11,"toc":415},"minimark",[12,25,30,33,36,44,47,50,54,60,66,69,85,92,96,99,110,113,189,192,196,203,206,306,309,313,326,329,332,336,357,360,363,367,406,409],[13,14,15,16,24],"p",{},"Until now, AI agents improved their answers. HyperAgents improves the mechanism that generates those improvements. It's the difference between learning to play chess better — and rewriting the rules of the game mid-match. Meta just published the ",[17,18,23],"a",{"href":19,"rel":20,"target":22},"https:\u002F\u002Farxiv.org\u002Fabs\u002F2603.19461",[21],"nofollow","\\_blank","arXiv paper 2603.19461"," and the code. It's open-source.",[26,27,29],"h2",{"id":28},"the-infinite-regress-problem","The Infinite Regress Problem",[13,31,32],{},"Current AI agents follow a simple pattern. An agent executes a task. A meta-agent — a separate system — observes the agent and improves its strategy. The meta-agent is the coach. The agent is the player.",[13,34,35],{},"The problem: who improves the coach?",[13,37,38,39,43],{},"To improve the meta-agent, you'd need a meta-meta-agent. To improve that one, a meta-meta-meta-agent. The chain never ends. This is called ",[40,41,42],"strong",{},"infinite regress"," — an endless regression where each layer of improvement requires an additional layer above it.",[13,45,46],{},"In practice, current architectures cut the chain arbitrarily. The meta-agent is frozen. It never improves. The agent gets better, but its capacity to get better stays fixed.",[13,48,49],{},"HyperAgents solves this by removing the separation entirely. The agent and meta-agent merge into a single self-referential program — a system that can modify its own code, including the rules that define how it improves.",[26,51,53],{"id":52},"metacognitive-self-modification-thinking-about-how-you-think","Metacognitive Self-Modification: Thinking About How You Think",[13,55,56,59],{},[40,57,58],{},"Metacognition"," is the ability to think about your own thinking. A simple human example: you know you learn better in the morning. So you reorganize your schedule to study early. That's not just learning. That's modifying your learning strategy.",[13,61,62,65],{},[40,63,64],{},"Metacognitive self-modification"," applies this principle to AI. A HyperAgents agent doesn't just solve a problem. It evaluates how it solves the problem. And it rewrites its own learning mechanism mid-task to solve the next ones better.",[13,67,68],{},"Concretely, the agent is an editable program — not a frozen black box. It can rewrite its own learning rules during execution. Two levels operate simultaneously within a single system:",[70,71,72,79],"ol",{},[73,74,75,78],"li",{},[40,76,77],{},"The task agent"," — solve the immediate problem",[73,80,81,84],{},[40,82,83],{},"The meta-agent"," — improve the learning strategy",[13,86,87,88,91],{},"Merged, these two levels create a ",[40,89,90],{},"self-referential"," program: a system that takes itself as its own object of improvement, without requiring an additional layer.",[26,93,95],{"id":94},"_00-0710-on-academic-paper-peer-review","0.0 → 0.710 on Academic Paper Peer Review",[13,97,98],{},"The results are spectacular. HyperAgents was tested across 4 radically different domains: code, academic paper peer review, robotics, and olympiad mathematics.",[13,100,101,102,105,106,109],{},"The most striking benchmark: ",[40,103,104],{},"peer review",". Classical agents score 0.0 on scientific paper quality evaluation. HyperAgents scores ",[40,107,108],{},"0.710"," — beating AI-Scientist-v2, currently the best AI system for automated research.",[13,111,112],{},"This isn't a marginal improvement. It's an order of magnitude.",[114,115,116,135],"table",{},[117,118,119],"thead",{},[120,121,122,126,129,132],"tr",{},[123,124,125],"th",{},"Domain",[123,127,128],{},"Baseline",[123,130,131],{},"AI-Scientist-v2",[123,133,134],{},"HyperAgents",[136,137,138,154,167,178],"tbody",{},[120,139,140,144,147,150],{},[141,142,143],"td",{},"Paper peer review",[141,145,146],{},"0.0",[141,148,149],{},"~0.5 (estimated)",[141,151,152],{},[40,153,108],{},[120,155,156,159,162,164],{},[141,157,158],{},"Code",[141,160,161],{},"—",[141,163,161],{},[141,165,166],{},"✅ SOTA",[120,168,169,172,174,176],{},[141,170,171],{},"Robotics",[141,173,161],{},[141,175,161],{},[141,177,166],{},[120,179,180,183,185,187],{},[141,181,182],{},"Olympiad math",[141,184,161],{},[141,186,161],{},[141,188,166],{},[13,190,191],{},"The logic is clear. If an agent can evaluate scientific papers with this accuracy, it can potentially write better ones. The line between research assistant and autonomous researcher just got thinner.",[26,193,195],{"id":194},"_4-domains-1-architecture","4 Domains, 1 Architecture",[13,197,198,199,202],{},"What makes HyperAgents fundamentally different: it's not a specialized model. It's a ",[40,200,201],{},"general architecture"," that adapts to radically different domains — code, scientific prose, robotic control, mathematical reasoning.",[13,204,205],{},"The key: self-modification allows the agent to adapt its strategy to the domain without being retrained for each new context. It discovers by itself how to learn in a domain it has never seen.",[114,207,208,223],{},[117,209,210],{},[120,211,212,215,218,221],{},[123,213,214],{},"Criteria",[123,216,217],{},"Classical agent",[123,219,220],{},"Agent + Meta-agent",[123,222,134],{},[136,224,225,237,250,263,278,291],{},[120,226,227,230,233,235],{},[141,228,229],{},"Improves its answers",[141,231,232],{},"✅",[141,234,232],{},[141,236,232],{},[120,238,239,242,245,248],{},[141,240,241],{},"Improves its strategy",[141,243,244],{},"❌",[141,246,247],{},"✅ partial",[141,249,232],{},[120,251,252,255,257,259],{},[141,253,254],{},"Improves its improvement mechanism",[141,256,244],{},[141,258,244],{},[141,260,261],{},[40,262,232],{},[120,264,265,268,270,273],{},[141,266,267],{},"Infinite regress",[141,269,161],{},[141,271,272],{},"✅ Problem",[141,274,275],{},[40,276,277],{},"❌ Solved",[120,279,280,283,285,287],{},[141,281,282],{},"Self-referential program",[141,284,244],{},[141,286,244],{},[141,288,289],{},[40,290,232],{},[120,292,293,296,299,301],{},[141,294,295],{},"Open-source",[141,297,298],{},"Varies",[141,300,298],{},[141,302,303],{},[40,304,305],{},"✅ GitHub",[13,307,308],{},"This is the fulfilled promise of reinforcement learning applied to agents: a system that not only learns, but learns to learn.",[26,310,312],{"id":311},"open-source-5-institutions-the-political-signal","Open-Source + 5 Institutions: The Political Signal",[13,314,315,316,319,320,325],{},"Meta chose to publish the paper ",[40,317,318],{},"and"," the code — available on ",[17,321,324],{"href":322,"rel":323,"target":22},"https:\u002F\u002Fhuggingface.co\u002Fpapers\u002F2603.19461",[21],"facebookresearch\u002FHyperAgents"," and listed on Hugging Face. With the University of British Columbia, the University of Edinburgh, NYU, and Meta Superintelligence Labs. Not just an internal lab effort.",[13,327,328],{},"This is a deliberate signal. Research on agent self-improvement stays open. At a time when OpenAI increasingly closes its models and labs debate the \"danger\" of self-modification, Meta publishes everything.",[13,330,331],{},"The contrast is stark. OpenAI keeps o3 and GPT-4.5 behind closed APIs. Anthropic limits Claude Computer Use to paying customers. And Meta open-sources a self-referential agent capable of metacognitive self-modification on GitHub.",[26,333,335],{"id":334},"the-week-everything-shifted","The Week Everything Shifted",[13,337,338,342,343,342,347,351,352,356],{},[17,339,341],{"href":340},"\u002Fnews\u002Ffigma-canvas-agents-mcp-server-skills-claude-codex-cursor","Figma opened its canvas to agents on Monday",". ",[17,344,346],{"href":345},"\u002Fnews\u002Flinear-agent-issue-tracking-dead-ai-agents-product-management","Linear declared issue tracking dead on Tuesday",[17,348,350],{"href":349},"\u002Fnews\u002Fai2-molmoweb-open-source-web-agent-autonomous-navigation-gpt4o","AI2 released MolmoWeb",", an open-source web agent that outperforms GPT-4o. ",[17,353,355],{"href":354},"\u002Fnews\u002Fnvidia-gtc-2026-keynote-jensen-huang","Jensen Huang declared \"we achieved AGI\""," at GTC.",[13,358,359],{},"And Meta just released an agent capable of rewriting its own learning rules.",[13,361,362],{},"Agents are no longer just improving at tasks. They're starting to improve at how they improve. HyperAgents gives Jensen Huang's declaration concrete technical resonance.",[26,364,366],{"id":365},"key-takeaways","Key Takeaways",[368,369,370,383,389,396,403],"ul",{},[73,371,372,375,376,378,379,382],{},[40,373,374],{},"Meta AI"," releases ",[40,377,134],{}," with UBC, Edinburgh, NYU and Meta Superintelligence Labs — ",[17,380,23],{"href":19,"rel":381,"target":22},[21],", open-source code on GitHub",[73,384,385,386,388],{},"HyperAgents solves the ",[40,387,42],{}," of agent + meta-agent architectures by merging both into a single self-referential editable program",[73,390,391,392,395],{},"Core concept: ",[40,393,394],{},"metacognitive self-modification"," — the agent can modify the mechanism that defines how it learns, mid-task",[73,397,398,399,402],{},"Benchmark results: peer review ",[40,400,401],{},"0.0 → 0.710"," (beats AI-Scientist-v2), state-of-the-art in code, robotics and olympiad math",[73,404,405],{},"First open-source system capable of metacognitive self-modification — available immediately on GitHub",[407,408],"hr",{},[13,410,411],{},[412,413,414],"em",{},"This week, Figma opened its canvas to agents, Linear declared issue tracking dead, MolmoWeb put a web agent in every developer's hands. And Meta just released an agent capable of rewriting its own learning rules. Agents are no longer improving at tasks. They're starting to improve at how they improve. That's different.",{"title":416,"searchDepth":417,"depth":417,"links":418},"",2,[419,420,421,422,423,424,425],{"id":28,"depth":417,"text":29},{"id":52,"depth":417,"text":53},{"id":94,"depth":417,"text":95},{"id":194,"depth":417,"text":195},{"id":311,"depth":417,"text":312},{"id":334,"depth":417,"text":335},{"id":365,"depth":417,"text":366},"ai","2026-03-26T09:10:00","Meta AI releases HyperAgents, a self-referential agent architecture capable of modifying its own learning mechanism. Peer review score: 0.0 → 0.710, beating AI-Scientist-v2. Open-source code available.","md","\u002Fmeta-hyperagents-self-modification-metacognitive-apprentissage-autonome.jpg",{},true,"\u002Fnews\u002Fmeta-hyperagents-self-modification-metacognitive-autonomous-learning",{"title":6,"description":428},"Meta AI releases HyperAgents — self-referential agents with metacognitive self-modification. Peer review score 0.0 → 0.710, beating AI-Scientist-v2. Open-source on GitHub.","Meta HyperAgents: AI With Metacognitive Self-Modification","news\u002Fmeta-hyperagents-self-modification-metacognitive-autonomous-learning",[439,440,441,442,443,444,445,446,447,448],"meta","hyperagents","self-modification","metacognition","ai-agents","reinforcement-learning","arxiv","open-source","ai-scientist","llm","meta-hyperagents-self-modification-metacognitive","xV3PQsGWXHZjU3kQRb9BNUMb0v7FADePg-_KbaNrHow",[452,458,465],{"title":453,"description":454,"path":455,"date":456,"image":457,"category":426},"Meta pauses its new ad and business terms after the AI scraping backlash","Meta has indefinitely paused updates to four sets of commercial and advertising terms that publishers said would let it crawl their sites to build AI and ad products.","\u002Fnews\u002Fmeta-pauses-business-ad-terms-ai-scraping","2026-11-09T08:00:00","\u002Fimages\u002Fnews\u002Fmeta-pauses-business-ad-terms-ai-scraping.jpg",{"title":459,"description":460,"path":461,"date":462,"image":463,"category":464},"Claude Code's “don’t ask again” left passwords in 1,122 public GitHub files","Security firm Apiiro found 1,122 public copies of Claude Code’s local settings file holding secrets, after “Yes, and don’t ask again” saved approved commands in plain text.","\u002Fnews\u002Fclaude-code-dont-ask-again-passwords-github","2026-11-08T08:00:00","\u002Fimages\u002Fnews\u002Fclaude-code-dont-ask-again-passwords-github.jpg","vibecoding",{"title":466,"description":467,"path":468,"date":469,"image":470,"category":464},"Cursor cloud agents now build, host and publish live apps","Cursor's cloud agents can scaffold a web app, provision Postgres and Cursor SSO, run security checks, then publish a shareable URL — without a separate hosting account.","\u002Fnews\u002Fcursor-live-apps-cloud-agents-deploy-postgres-auth","2026-11-07T08:00:00","\u002Fimages\u002Fnews\u002Fcursor-live-apps-cloud-agents-deploy-postgres-auth.jpg",{"en":472,"fr":760,"es":8},{"id":5,"title":6,"author":7,"author_avatar":8,"body":473,"category":426,"date":427,"description":428,"extension":429,"image":430,"meta":757,"navigation":432,"path":433,"seo":758,"seoDescription":435,"seoTitle":436,"stem":437,"tags":759,"translationKey":449,"updatedAt":8,"__hash__":450},{"type":10,"value":474,"toc":748},[475,480,482,484,486,490,492,494,496,500,504,506,516,520,522,524,530,532,590,592,594,598,600,684,686,688,695,697,699,701,711,713,715,717,742,744],[13,476,15,477,24],{},[17,478,23],{"href":19,"rel":479,"target":22},[21],[26,481,29],{"id":28},[13,483,32],{},[13,485,35],{},[13,487,38,488,43],{},[40,489,42],{},[13,491,46],{},[13,493,49],{},[26,495,53],{"id":52},[13,497,498,59],{},[40,499,58],{},[13,501,502,65],{},[40,503,64],{},[13,505,68],{},[70,507,508,512],{},[73,509,510,78],{},[40,511,77],{},[73,513,514,84],{},[40,515,83],{},[13,517,87,518,91],{},[40,519,90],{},[26,521,95],{"id":94},[13,523,98],{},[13,525,101,526,105,528,109],{},[40,527,104],{},[40,529,108],{},[13,531,112],{},[114,533,534,546],{},[117,535,536],{},[120,537,538,540,542,544],{},[123,539,125],{},[123,541,128],{},[123,543,131],{},[123,545,134],{},[136,547,548,560,570,580],{},[120,549,550,552,554,556],{},[141,551,143],{},[141,553,146],{},[141,555,149],{},[141,557,558],{},[40,559,108],{},[120,561,562,564,566,568],{},[141,563,158],{},[141,565,161],{},[141,567,161],{},[141,569,166],{},[120,571,572,574,576,578],{},[141,573,171],{},[141,575,161],{},[141,577,161],{},[141,579,166],{},[120,581,582,584,586,588],{},[141,583,182],{},[141,585,161],{},[141,587,161],{},[141,589,166],{},[13,591,191],{},[26,593,195],{"id":194},[13,595,198,596,202],{},[40,597,201],{},[13,599,205],{},[114,601,602,614],{},[117,603,604],{},[120,605,606,608,610,612],{},[123,607,214],{},[123,609,217],{},[123,611,220],{},[123,613,134],{},[136,615,616,626,636,648,660,672],{},[120,617,618,620,622,624],{},[141,619,229],{},[141,621,232],{},[141,623,232],{},[141,625,232],{},[120,627,628,630,632,634],{},[141,629,241],{},[141,631,244],{},[141,633,247],{},[141,635,232],{},[120,637,638,640,642,644],{},[141,639,254],{},[141,641,244],{},[141,643,244],{},[141,645,646],{},[40,647,232],{},[120,649,650,652,654,656],{},[141,651,267],{},[141,653,161],{},[141,655,272],{},[141,657,658],{},[40,659,277],{},[120,661,662,664,666,668],{},[141,663,282],{},[141,665,244],{},[141,667,244],{},[141,669,670],{},[40,671,232],{},[120,673,674,676,678,680],{},[141,675,295],{},[141,677,298],{},[141,679,298],{},[141,681,682],{},[40,683,305],{},[13,685,308],{},[26,687,312],{"id":311},[13,689,315,690,319,692,325],{},[40,691,318],{},[17,693,324],{"href":322,"rel":694,"target":22},[21],[13,696,328],{},[13,698,331],{},[26,700,335],{"id":334},[13,702,703,342,705,342,707,351,709,356],{},[17,704,341],{"href":340},[17,706,346],{"href":345},[17,708,350],{"href":349},[17,710,355],{"href":354},[13,712,359],{},[13,714,362],{},[26,716,366],{"id":365},[368,718,719,728,732,736,740],{},[73,720,721,375,723,378,725,382],{},[40,722,374],{},[40,724,134],{},[17,726,23],{"href":19,"rel":727,"target":22},[21],[73,729,385,730,388],{},[40,731,42],{},[73,733,391,734,395],{},[40,735,394],{},[73,737,398,738,402],{},[40,739,401],{},[73,741,405],{},[407,743],{},[13,745,746],{},[412,747,414],{},{"title":416,"searchDepth":417,"depth":417,"links":749},[750,751,752,753,754,755,756],{"id":28,"depth":417,"text":29},{"id":52,"depth":417,"text":53},{"id":94,"depth":417,"text":95},{"id":194,"depth":417,"text":195},{"id":311,"depth":417,"text":312},{"id":334,"depth":417,"text":335},{"id":365,"depth":417,"text":366},{},{"title":6,"description":428},[439,440,441,442,443,444,445,446,447,448],{"id":761,"title":762,"author":7,"author_avatar":8,"body":763,"category":426,"date":427,"description":1138,"extension":429,"image":430,"meta":1139,"navigation":432,"path":1140,"seo":1141,"seoDescription":1142,"seoTitle":1143,"stem":1144,"tags":1145,"translationKey":449,"updatedAt":8,"__hash__":1147},"news_fr\u002Ffr\u002Fnews\u002Fmeta-hyperagents-self-modification-metacognitive-apprentissage-autonome.md","Meta publie HyperAgents : l'IA qui améliore sa propre façon de s'améliorer — la self-modification metacognitive expliquée",{"type":10,"value":764,"toc":1129},[765,773,777,780,783,789,792,795,799,806,812,815,829,836,840,843,852,855,918,921,925,932,935,1029,1032,1036,1047,1050,1053,1057,1077,1080,1083,1087,1122,1124],[13,766,767,768,772],{},"Jusqu'ici, les agents IA amélioraient leurs réponses. HyperAgents améliore le mécanisme qui génère ces améliorations. C'est la différence entre apprendre à mieux jouer aux échecs — et réécrire les règles du jeu pendant la partie. Meta vient de publier le ",[17,769,771],{"href":19,"rel":770,"target":22},[21],"paper arXiv 2603.19461"," et le code. C'est open-source.",[26,774,776],{"id":775},"le-problème-de-linfinite-regress","Le problème de l'infinite regress",[13,778,779],{},"Les agents IA actuels fonctionnent sur un schéma simple. Un agent exécute une tâche. Un meta-agent — un second système — observe l'agent et améliore sa stratégie. Le meta-agent est le coach. L'agent est le joueur.",[13,781,782],{},"Le problème : qui améliore le coach ?",[13,784,785,786,788],{},"Pour améliorer le meta-agent, il faudrait un méta-méta-agent. Pour améliorer celui-ci, un méta-méta-méta-agent. La chaîne ne s'arrête jamais. C'est ce qu'on appelle l'",[40,787,42],{}," — une régression à l'infini où chaque couche d'amélioration nécessite une couche supplémentaire au-dessus.",[13,790,791],{},"En pratique, les architectures actuelles coupent la chaîne arbitrairement. Le meta-agent est figé. Il ne s'améliore jamais. L'agent progresse, mais sa capacité à progresser reste fixe.",[13,793,794],{},"HyperAgents résout ce problème en supprimant la séparation. L'agent et le meta-agent fusionnent en un seul programme auto-référentiel — un système qui peut modifier son propre code, y compris les règles qui définissent comment il s'améliore.",[26,796,798],{"id":797},"la-metacognitive-self-modification-penser-à-comment-on-pense","La metacognitive self-modification : penser à comment on pense",[13,800,801,802,805],{},"La ",[40,803,804],{},"métacognition",", c'est la capacité de penser à sa propre façon de penser. Chez l'humain, un exemple simple : tu sais que tu apprends mieux le matin. Donc tu réorganises ton emploi du temps pour étudier tôt. Ce n'est pas juste apprendre. C'est modifier ta stratégie d'apprentissage.",[13,807,801,808,811],{},[40,809,810],{},"self-modification metacognitive"," applique ce principe à une IA. L'agent HyperAgents ne se contente pas de résoudre un problème. Il évalue comment il résout le problème. Et il réécrit son propre mécanisme d'apprentissage en cours de tâche pour mieux résoudre les prochains.",[13,813,814],{},"Concrètement, l'agent est un programme éditable — pas une boîte noire figée. Il peut réécrire ses propres règles d'apprentissage pendant l'exécution. Deux niveaux fonctionnent simultanément dans un seul système :",[70,816,817,823],{},[73,818,819,822],{},[40,820,821],{},"L'agent de tâche"," — résoudre le problème immédiat",[73,824,825,828],{},[40,826,827],{},"Le meta-agent"," — améliorer la stratégie d'apprentissage",[13,830,831,832,835],{},"Fusionnés, ces deux niveaux créent un programme ",[40,833,834],{},"auto-référentiel"," : un système qui se prend lui-même comme objet d'amélioration, sans nécessiter de couche supplémentaire.",[26,837,839],{"id":838},"_00-0710-sur-le-peer-review-de-papiers-académiques","0.0 → 0.710 sur le peer review de papiers académiques",[13,841,842],{},"Les résultats sont spectaculaires. HyperAgents a été testé sur 4 domaines radicalement différents : code, peer review de papiers scientifiques, robotique, et mathématiques olympiades.",[13,844,845,846,848,849,851],{},"Le benchmark le plus frappant : le ",[40,847,104],{},". Les agents classiques obtiennent un score de 0.0 sur l'évaluation de qualité de papers scientifiques. HyperAgents obtient ",[40,850,108],{}," — battant AI-Scientist-v2, actuellement le meilleur système d'IA pour la recherche automatisée.",[13,853,854],{},"Ce n'est pas une amélioration marginale. C'est un ordre de grandeur.",[114,856,857,870],{},[117,858,859],{},[120,860,861,864,866,868],{},[123,862,863],{},"Domaine",[123,865,128],{},[123,867,131],{},[123,869,134],{},[136,871,872,886,896,907],{},[120,873,874,877,879,882],{},[141,875,876],{},"Peer review papiers",[141,878,146],{},[141,880,881],{},"~0.5 (estimé)",[141,883,884],{},[40,885,108],{},[120,887,888,890,892,894],{},[141,889,158],{},[141,891,161],{},[141,893,161],{},[141,895,166],{},[120,897,898,901,903,905],{},[141,899,900],{},"Robotique",[141,902,161],{},[141,904,161],{},[141,906,166],{},[120,908,909,912,914,916],{},[141,910,911],{},"Maths olympiades",[141,913,161],{},[141,915,161],{},[141,917,166],{},[13,919,920],{},"La logique est implacable. Si un agent peut évaluer des papers scientifiques avec cette précision, il peut potentiellement en écrire de meilleurs. La frontière entre assistant de recherche et chercheur autonome vient de se réduire.",[26,922,924],{"id":923},"_4-domaines-1-architecture","4 domaines, 1 architecture",[13,926,927,928,931],{},"Ce qui rend HyperAgents fondamentalement différent : ce n'est pas un modèle spécialisé. C'est une ",[40,929,930],{},"architecture générale"," qui s'adapte à des domaines radicalement différents — code, prose scientifique, contrôle robotique, raisonnement mathématique.",[13,933,934],{},"La clé : la self-modification permet à l'agent d'adapter sa stratégie au domaine sans être ré-entraîné pour chaque nouveau contexte. Il découvre lui-même comment apprendre dans un domaine qu'il n'a jamais vu.",[114,936,937,951],{},[117,938,939],{},[120,940,941,944,947,949],{},[123,942,943],{},"Critère",[123,945,946],{},"Agent classique",[123,948,220],{},[123,950,134],{},[136,952,953,964,976,989,1003,1016],{},[120,954,955,958,960,962],{},[141,956,957],{},"Améliore ses réponses",[141,959,232],{},[141,961,232],{},[141,963,232],{},[120,965,966,969,971,974],{},[141,967,968],{},"Améliore sa stratégie",[141,970,244],{},[141,972,973],{},"✅ partiel",[141,975,232],{},[120,977,978,981,983,985],{},[141,979,980],{},"Améliore son mécanisme d'amélioration",[141,982,244],{},[141,984,244],{},[141,986,987],{},[40,988,232],{},[120,990,991,993,995,998],{},[141,992,267],{},[141,994,161],{},[141,996,997],{},"✅ Problème",[141,999,1000],{},[40,1001,1002],{},"❌ Résolu",[120,1004,1005,1008,1010,1012],{},[141,1006,1007],{},"Programme auto-référentiel",[141,1009,244],{},[141,1011,244],{},[141,1013,1014],{},[40,1015,232],{},[120,1017,1018,1020,1023,1025],{},[141,1019,295],{},[141,1021,1022],{},"Variable",[141,1024,1022],{},[141,1026,1027],{},[40,1028,305],{},[13,1030,1031],{},"C'est la promesse tenue du reinforcement learning appliqué aux agents : un système qui non seulement apprend, mais apprend à apprendre.",[26,1033,1035],{"id":1034},"open-source-5-institutions-le-signal-politique","Open-source + 5 institutions : le signal politique",[13,1037,1038,1039,1042,1043,1046],{},"Meta choisit de publier le paper ",[40,1040,1041],{},"et"," le code — disponible sur ",[17,1044,324],{"href":322,"rel":1045,"target":22},[21]," et répertorié sur Hugging Face. Avec l'University of British Columbia, l'University of Edinburgh, NYU et Meta Superintelligence Labs. Pas seulement un lab interne.",[13,1048,1049],{},"C'est un signal délibéré. La recherche sur l'auto-amélioration des agents reste ouverte. À l'heure où OpenAI ferme de plus en plus ses modèles et où les labs débattent de la \"dangerosité\" de la self-modification, Meta publie tout.",[13,1051,1052],{},"Le contraste est net. OpenAI garde o3 et GPT-4.5 derrière des APIs fermées. Anthropic limite Claude Computer Use à ses clients payants. Et Meta met un agent auto-référentiel capable de self-modification metacognitive en open-source sur GitHub.",[26,1054,1056],{"id":1055},"la-semaine-où-tout-a-basculé","La semaine où tout a basculé",[13,1058,1059,342,1063,342,1067,1071,1072,1076],{},[17,1060,1062],{"href":1061},"\u002Ffr\u002Fnews\u002Ffigma-canvas-agents-mcp-server-skills-claude-codex-cursor","Figma a ouvert son canvas aux agents lundi",[17,1064,1066],{"href":1065},"\u002Ffr\u002Fnews\u002Flinear-agent-issue-tracking-mort-agents-ia-product-management","Linear a déclaré la mort du issue tracking mardi",[17,1068,1070],{"href":1069},"\u002Ffr\u002Fnews\u002Fai2-molmoweb-agent-web-open-source-navigation-autonome-gpt4o","AI2 a publié MolmoWeb",", un agent web open-source qui surpasse GPT-4o. ",[17,1073,1075],{"href":1074},"\u002Ffr\u002Fnews\u002Fnvidia-gtc-2026-keynote-jensen-huang","Jensen Huang a déclaré \"we achieved AGI\""," lors de la GTC.",[13,1078,1079],{},"Et Meta vient de publier un agent capable de réécrire ses propres règles d'apprentissage.",[13,1081,1082],{},"Les agents ne s'améliorent plus seulement sur les tâches. Ils commencent à s'améliorer sur la façon dont ils s'améliorent. HyperAgents donne à la déclaration de Jensen Huang une résonance technique concrète.",[26,1084,1086],{"id":1085},"en-résumé","En résumé",[368,1088,1089,1101,1107,1113,1119],{},[73,1090,1091,1093,1094,1096,1097,1100],{},[40,1092,374],{}," publie ",[40,1095,134],{}," avec UBC, Edinburgh, NYU et Meta Superintelligence Labs — ",[17,1098,771],{"href":19,"rel":1099,"target":22},[21],", code open-source sur GitHub",[73,1102,1103,1104,1106],{},"HyperAgents résout l'",[40,1105,42],{}," des architectures agent + meta-agent en fusionnant les deux en un seul programme auto-référentiel éditable",[73,1108,1109,1110,1112],{},"Concept clé : ",[40,1111,394],{}," — l'agent peut modifier le mécanisme qui définit comment il apprend, en cours de tâche",[73,1114,1115,1116,1118],{},"Résultats benchmarks : peer review ",[40,1117,401],{}," (bat AI-Scientist-v2), performances état de l'art en code, robotique et maths olympiades",[73,1120,1121],{},"Premier système open-source capable de self-modification metacognitive — disponible immédiatement sur GitHub",[407,1123],{},[13,1125,1126],{},[412,1127,1128],{},"Cette semaine, Figma a ouvert son canvas aux agents, Linear a déclaré la mort du issue tracking, MolmoWeb a mis un agent web dans les mains de n'importe quel développeur. Et Meta vient de publier un agent capable de réécrire ses propres règles d'apprentissage. Les agents ne s'améliorent plus sur les tâches. Ils commencent à s'améliorer sur la façon dont ils s'améliorent. C'est différent.",{"title":416,"searchDepth":417,"depth":417,"links":1130},[1131,1132,1133,1134,1135,1136,1137],{"id":775,"depth":417,"text":776},{"id":797,"depth":417,"text":798},{"id":838,"depth":417,"text":839},{"id":923,"depth":417,"text":924},{"id":1034,"depth":417,"text":1035},{"id":1055,"depth":417,"text":1056},{"id":1085,"depth":417,"text":1086},"Meta AI publie HyperAgents, une architecture d'agents auto-référentiels capables de modifier leur propre mécanisme d'apprentissage. Score peer review : 0.0 → 0.710, battant AI-Scientist-v2. Code open-source.",{},"\u002Ffr\u002Fnews\u002Fmeta-hyperagents-self-modification-metacognitive-apprentissage-autonome",{"title":762,"description":1138},"Meta AI publie HyperAgents — agents auto-référentiels à self-modification metacognitive. Score peer review 0.0 → 0.710, battant AI-Scientist-v2. Open-source sur GitHub.","HyperAgents : Meta publie l'IA à self-modification metacognitive","fr\u002Fnews\u002Fmeta-hyperagents-self-modification-metacognitive-apprentissage-autonome",[439,440,441,442,1146,444,445,446,447,448],"agents-ia","GXUp05gOXis8lWgr8pTM5gu_i4ohZ_bFwwM2TOmCIIQ"]