{"id":15998,"date":"2026-11-09T13:16:23","date_gmt":"2026-11-09T12:16:23","guid":{"rendered":"https:\/\/dailyai.com\/?p=15998"},"modified":"2026-11-09T13:16:23","modified_gmt":"2026-11-09T12:16:23","slug":"best-coding-agent-fully-clears-only-26-4-of-real-tasks","status":"publish","type":"post","link":"https:\/\/dailyai.com\/fr\/2026\/11\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\/","title":{"rendered":"Le meilleur agent de codage ne r\u00e9ussit que 26,4 % des t\u00e2ches r\u00e9elles."},"content":{"rendered":"

26,4 %. C'est la proportion de t\u00e2ches d'\u00e9dition d'applications r\u00e9alis\u00e9es par de v\u00e9ritables utilisateurs que l'agent de codage le plus performant a enti\u00e8rement valid\u00e9es sur la nouvelle plateforme de co-conception d'Alibaba, selon une \u00e9tude d'Alibaba Token Hub et de ses co-auteurs universitaires. Lors de cette m\u00eame ex\u00e9cution, 69,9 % des crit\u00e8res d'acceptation implicites de chaque t\u00e2che ont \u00e9t\u00e9 atteints en moyenne. La plupart des t\u00e2ches \u00e9taient presque termin\u00e9es, mais restaient encore inachev\u00e9es.<\/strong><\/p>\n\n\n\n

Les Document de r\u00e9f\u00e9rence sur la co-conception<\/a> (arXiv, 26 octobre 2026\u00a0; site du projet \u00e0 l'adresse codesignbench.github.io<\/a>Ce syst\u00e8me g\u00e9n\u00e8re 273 t\u00e2ches \u00e0 partir de 37\u00a0463 requ\u00eates d\u00e9dupliqu\u00e9es, r\u00e9parties sur 26\u00a0314 sessions de d\u00e9veloppement r\u00e9elles. Les agents doivent modifier une application web fonctionnelle \u00e0 partir d'une requ\u00eate utilisateur fixe, incluant parfois des pi\u00e8ces jointes (images, fichiers audio, vid\u00e9os ou documents). Des experts ont r\u00e9dig\u00e9 2\u00a0331 crit\u00e8res d'acceptation cach\u00e9s. Un premier syst\u00e8me d'\u00e9valuation par IA note l'application finale \u00e0 partir du code source et des donn\u00e9es du navigateur\u00a0; un second syst\u00e8me v\u00e9rifie chaque verdict \u00ab\u00a0R\u00e9ussi\u00a0\u00bb ou \u00ab\u00a0Partiel\u00a0\u00bb et recherche les \u00e9bauches, le code cod\u00e9 en dur et les fonctionnalit\u00e9s inaccessibles.<\/p>\n\n\n\n

Que signifient les chiffres principaux ?<\/h2>\n\n\n\n

Le mod\u00e8le GPT-6 Astra d'OpenAI a \u00e9t\u00e9 test\u00e9 sur dix points d'ex\u00e9cution avec l'environnement Claude Code. Sur l'ensemble des 273 t\u00e2ches, il a obtenu un score de r\u00e9ussite stricte moyen de 0,699 (environ sept crit\u00e8res sur dix r\u00e9ussis par t\u00e2che) et un taux de r\u00e9ussite compl\u00e8te de 26,4 %. En d'autres termes, dans 73,6 % des t\u00e2ches, au moins un crit\u00e8re n'a pas \u00e9t\u00e9 respect\u00e9. La couverture moyenne des exigences ne signifie pas la mise en \u0153uvre d'une modification compl\u00e8te.<\/p>\n\n\n\n

L'\u00e9cart est plus marqu\u00e9 selon le type de t\u00e2che. Sur 100 t\u00e2ches d'\u00e9dition multimodale, Astra a r\u00e9ussi 41 % des t\u00e2ches, avec un taux de r\u00e9ussite stricte moyen de 0,754. Sur 173 t\u00e2ches d'\u00e9dition de r\u00e9f\u00e9rentiel (modifications coordonn\u00e9es entre composants, routes, \u00e9tat et flux de donn\u00e9es), le taux de r\u00e9ussite stricte s'est \u00e9tabli \u00e0 17,9 %, avec un taux de 0,667. Les modifications de r\u00e9f\u00e9rentiel portant sur trois ou quatre types de modifications ont obtenu des scores inf\u00e9rieurs pour chaque point de terminaison par rapport \u00e0 celles portant sur un ou deux types. Le taux de r\u00e9ussite stricte cumul\u00e9 est pass\u00e9 de 0,654 pour les modifications plus cibl\u00e9es \u00e0 0,591 pour les modifications plus g\u00e9n\u00e9rales.<\/p>\n\n\n\n

Ce mod\u00e8le correspond \u00e0 la mani\u00e8re dont les utilisateurs con\u00e7oivent r\u00e9ellement les outils de d\u00e9veloppement. Une requ\u00eate qui modifie une seule propri\u00e9t\u00e9 CSS est diff\u00e9rente d'une t\u00e2che qui refond la navigation, l'\u00e9tat et un r\u00e9sum\u00e9 d\u00e9pendant en une seule \u00e9tape. Le parcours de test des d\u00e9p\u00f4ts est con\u00e7u pour que ces requ\u00eates complexes soient fr\u00e9quentes\u00a0: 166 des 173 t\u00e2ches de d\u00e9p\u00f4t n\u00e9cessitent plusieurs types de modifications.<\/p>\n\n\n\n

Pourquoi les anciens tests semblent trop faciles<\/h2>\n\n\n\n

Les auteurs ont con\u00e7u cet ensemble de tests car les anciens benchmarks d'\u00e9dition atteignent leurs limites d\u00e8s que les mod\u00e8les sont capables de boucler, d'ex\u00e9cuter et de corriger leur propre code. Sur 150 t\u00e2ches d'\u00e9dition WebCompass, en maintenant constants la grille d'\u00e9valuation et le r\u00f4le du juge, Qwen 3.7-Plus est pass\u00e9 d'un taux moyen de compl\u00e9tion des fonctionnalit\u00e9s not\u00e9es de 73,6 % en une seule ex\u00e9cution \u00e0 95,1 % avec une ex\u00e9cution automatis\u00e9e\u00a0; Claude Opus 4.8 a atteint 99,4 % avec une ex\u00e9cution automatis\u00e9e sur le m\u00eame ensemble de tests. De tels scores ne permettent plus de d\u00e9partager les produits.<\/p>\n\n\n\n

Co-Design Bench a privil\u00e9gi\u00e9 les demandes de session pr\u00e9sentant d\u00e9j\u00e0 des points de friction (reformulations, suivis ou annulations) et n'\u00e9value que les fonctionnalit\u00e9s op\u00e9rationnelles dans l'application en cours d'ex\u00e9cution. Les t\u00e2ches multimodales peuvent demander \u00e0 un agent d'extraire une valeur d'un re\u00e7u ou d'un extrait vid\u00e9o tout en assurant la coh\u00e9rence des totaux et des vues d\u00e9pendantes. Les t\u00e2ches de r\u00e9f\u00e9rentiel peuvent associer des limites de nombre d'extraits, la pr\u00e9cision de la chronologie et la r\u00e9cup\u00e9ration apr\u00e8s une s\u00e9paration. Un correctif local, m\u00eame plausible, peut n\u00e9anmoins laisser une vue li\u00e9e incorrecte.<\/p>\n\n\n\n

Qui a dirig\u00e9, et qui n'a pas dirig\u00e9<\/h2>\n\n\n\n

Derri\u00e8re Astra, Claude Opus 5 d'Anthropic et Qwen3.8-Max d'Alibaba affichaient des taux de r\u00e9ussite complets combin\u00e9s de 18,3 % et 16,5 % respectivement. Grok 4.6 de xAI fermait la marche avec 6,6 % de r\u00e9ussites compl\u00e8tes. Le classement variait selon les types de tests\u00a0: Qwen3.8-27B se hissait \u00e0 la deuxi\u00e8me place pour les modifications de d\u00e9p\u00f4t, malgr\u00e9 un classement inf\u00e9rieur pour les tests multimodaux. L'article n'interpr\u00e8te pas cela comme un avantage g\u00e9n\u00e9ral du mod\u00e8le, mais simplement comme la preuve qu'un seul tableau combin\u00e9 peut masquer des points forts sp\u00e9cifiques \u00e0 un contexte.<\/p>\n\n\n\n

L'utilisation accrue de jetons n'a pas permis de corriger le probl\u00e8me. Lors de l'ex\u00e9cution multimodale, Astra a atteint un taux de r\u00e9ussite stricte de 0,754 avec une m\u00e9diane de 32 tours, tandis que GLM-5.2 a utilis\u00e9 69 tours et un trafic de jetons plusieurs fois sup\u00e9rieur pour un score inf\u00e9rieur. Des ressources de calcul suppl\u00e9mentaires ont permis \u00e0 certaines ex\u00e9cutions de se prolonger sans que les crit\u00e8res inachev\u00e9s, bloquant ainsi une passe compl\u00e8te, ne soient trait\u00e9s.<\/p>\n\n\n\n

Le harnais d\u00e9place le tableau d'affichage<\/h2>\n\n\n\n

Lorsque l'\u00e9quipe a fix\u00e9 deux solveurs et interverti les outils d'analyse (Codex, OpenCode ou Qwen Code), les scores ont vari\u00e9 jusqu'\u00e0 16,2 points de pourcentage lors des modifications de d\u00e9p\u00f4ts. Le mod\u00e8le le plus performant d\u00e9pendait \u00e9galement de l'outil utilis\u00e9. Lors de modifications multimodales, le classement des outils s'est m\u00eame invers\u00e9 entre les solveurs. La conclusion de l'article est sans \u00e9quivoque\u00a0: il faut pr\u00e9senter les r\u00e9sultats conjointement avec le mod\u00e8le. Un classement qui ne mentionne que le mod\u00e8le ne donne qu'une vision partielle de la situation.<\/p>\n\n\n\n

Pour une entreprise qui choisit un agent de codage, cela signifie tester la paire que vous allez r\u00e9ellement ex\u00e9cuter \u2014 mod\u00e8le et outils \u2014 sur des modifications qui ressemblent \u00e0 votre backlog, et non pas se fier \u00e0 un graphique qui r\u00e9pertorie uniquement le mod\u00e8le.<\/p>\n\n\n\n

Ce que le test ne pr\u00e9tend pas<\/h2>\n\n\n\n

Les limites sont \u00e9nonc\u00e9es dans l'article. Chaque t\u00e2che correspond \u00e0 une requ\u00eate fixe\u00a0; l'agent ne peut ni poser de questions compl\u00e9mentaires ni recevoir de corrections en cours de t\u00e2che. Les \u00e9valuateurs IA peuvent encore commettre des erreurs, malgr\u00e9 la v\u00e9rification des scores par deux experts en d\u00e9veloppement front-end sur 200 r\u00e9sultats de mod\u00e8les de t\u00e2ches. Le r\u00e9sultat mesure l'ex\u00e9cution autonome selon un crit\u00e8re strict, et non le travail interactif avec une personne. Il ne refl\u00e8te pas non plus la fr\u00e9quence \u00e0 laquelle les humains corrigent le code courant. L'interpr\u00e9tation est plus nuanc\u00e9e\u00a0: lors de r\u00e9visions complexes en conditions r\u00e9elles, m\u00eame le meilleur r\u00e9sultat a souvent laiss\u00e9 des \u00e9l\u00e9ments inachev\u00e9s.<\/p>\n\n\n\n

Cette mise en garde rejoint une tendance que DailyAI a d\u00e9j\u00e0 observ\u00e9e ailleurs. METR a fait valoir que les vitesses annonc\u00e9es peuvent d\u00e9passer les preuves que les gens peuvent v\u00e9rifier.<\/a>, et les propres agents de s\u00e9curit\u00e9 d'OpenAI r\u00e9cemment Une \u00e9tude payante a permis de d\u00e9couvrir 948 failles d'\u00e9l\u00e9vation de privil\u00e8ges sous Linux.<\/a> \u2014 un rappel que les notes attribu\u00e9es aux agents et les r\u00e9sultats r\u00e9els sont des mesures diff\u00e9rentes.<\/p>\n\n\n\n

Pourquoi c'est important<\/h2>\n\n\n\n

Les fournisseurs continuent de vendre des agents de codage affichant des scores quasi parfaits sur d'anciens tests. Sur cet ensemble de tests r\u00e9els, le meilleur agent a laiss\u00e9 des \u00e9l\u00e9ments inachev\u00e9s dans environ trois t\u00e2ches sur quatre. Toute personne achetant ou d\u00e9ployant un agent devrait exiger des taux de r\u00e9ussite complets sur des modifications r\u00e9alistes, et pour le contexte d'utilisation de ces r\u00e9sultats, et non pas seulement une couverture moyenne des crit\u00e8res.<\/p>","protected":false},"excerpt":{"rendered":"

Sur la plateforme de co-conception d'Alibaba, l'agent de codage le plus performant a enti\u00e8rement ex\u00e9cut\u00e9 26,4 % des 273 t\u00e2ches d'\u00e9dition d'applications r\u00e9alis\u00e9es par de v\u00e9ritables utilisateurs. Il a n\u00e9anmoins satisfait en moyenne \u00e0 69,9 % des crit\u00e8res de chaque t\u00e2che\u00a0\u2014\u00a0la plupart du temps effectu\u00e9e, mais rarement termin\u00e9e.<\/p>","protected":false},"author":16,"featured_media":15999,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[],"tags":[],"class_list":["post-15998","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry"],"yoast_head":"\nCoding Agents Fully Clear Only 26.4% of Real Edits<\/title>\n<meta name=\"description\" content=\"Alibaba\u2019s Co-Design Bench finds the best coding agent fully clears only 26.4% of 273 real app-edit tasks, despite meeting 69.9% of criteria on average.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/dailyai.com\/fr\/2026\/11\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\/\" \/>\n<meta property=\"og:locale\" content=\"fr_FR\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Coding Agents Fully Clear Only 26.4% of Real Edits\" \/>\n<meta property=\"og:description\" content=\"Alibaba\u2019s Co-Design Bench finds the best coding agent fully clears only 26.4% of 273 real app-edit tasks, despite meeting 69.9% of criteria on average.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/dailyai.com\/fr\/2026\/11\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\/\" \/>\n<meta property=\"og:site_name\" content=\"DailyAI\" \/>\n<meta property=\"article:published_time\" content=\"2026-11-09T12:16:23+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/dailyai.com\/wp-content\/uploads\/2026\/10\/dailyai-og-default.jpg\" \/>\n\t<meta property=\"og:image:width\" content=\"1200\" \/>\n\t<meta property=\"og:image:height\" content=\"630\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/jpeg\" \/>\n<meta name=\"author\" content=\"David Merry\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:creator\" content=\"@DailyAIOfficial\" \/>\n<meta name=\"twitter:site\" content=\"@DailyAIOfficial\" \/>\n<meta name=\"twitter:label1\" content=\"\u00c9crit par\" \/>\n\t<meta name=\"twitter:data1\" content=\"David Merry\" \/>\n\t<meta name=\"twitter:label2\" content=\"Dur\u00e9e de lecture estim\u00e9e\" \/>\n\t<meta name=\"twitter:data2\" content=\"4 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"NewsArticle\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/2026\\\/11\\\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/2026\\\/11\\\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\\\/\"},\"author\":{\"name\":\"David Merry\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/#\\\/schema\\\/person\\\/285a97adb64b88751ac3440ece6b08a1\"},\"headline\":\"Best Coding Agent Fully Clears Only 26.4% of Real Tasks\",\"datePublished\":\"2026-11-09T12:16:23+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/2026\\\/11\\\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\\\/\"},\"wordCount\":906,\"publisher\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/#organization\"},\"image\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/2026\\\/11\\\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/dailyai.com\\\/wp-content\\\/uploads\\\/2026\\\/11\\\/dai-house-15998.jpg\",\"inLanguage\":\"fr-FR\"},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/2026\\\/11\\\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\\\/\",\"url\":\"https:\\\/\\\/dailyai.com\\\/2026\\\/11\\\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\\\/\",\"name\":\"Coding Agents Fully Clear Only 26.4% of Real Edits\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/2026\\\/11\\\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\\\/#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/2026\\\/11\\\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/dailyai.com\\\/wp-content\\\/uploads\\\/2026\\\/11\\\/dai-house-15998.jpg\",\"datePublished\":\"2026-11-09T12:16:23+00:00\",\"description\":\"Alibaba\u2019s Co-Design Bench finds the best coding agent fully clears only 26.4% of 273 real app-edit tasks, despite meeting 69.9% of criteria on average.\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/2026\\\/11\\\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\\\/#breadcrumb\"},\"inLanguage\":\"fr-FR\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/dailyai.com\\\/2026\\\/11\\\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\\\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"fr-FR\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/2026\\\/11\\\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\\\/#primaryimage\",\"url\":\"https:\\\/\\\/dailyai.com\\\/wp-content\\\/uploads\\\/2026\\\/11\\\/dai-house-15998.jpg\",\"contentUrl\":\"https:\\\/\\\/dailyai.com\\\/wp-content\\\/uploads\\\/2026\\\/11\\\/dai-house-15998.jpg\",\"width\":1472,\"height\":832,\"caption\":\"Flat vector illustration of a cream clipboard with three navy checklist bars and one unfinished red bar beside a navy code window.\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/2026\\\/11\\\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/dailyai.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Best Coding Agent Fully Clears Only 26.4% of Real Tasks\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/#website\",\"url\":\"https:\\\/\\\/dailyai.com\\\/\",\"name\":\"DailyAI\",\"description\":\"The AI economy, measured.\",\"publisher\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/#organization\"},\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/dailyai.com\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"fr-FR\"},{\"@type\":\"Organization\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/#organization\",\"name\":\"DailyAI\",\"url\":\"https:\\\/\\\/dailyai.com\\\/\",\"logo\":{\"@type\":\"ImageObject\",\"inLanguage\":\"fr-FR\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/#\\\/schema\\\/logo\\\/image\\\/\",\"url\":\"https:\\\/\\\/dailyai.com\\\/wp-content\\\/uploads\\\/2023\\\/06\\\/Daily-Ai_TL_colour.png\",\"contentUrl\":\"https:\\\/\\\/dailyai.com\\\/wp-content\\\/uploads\\\/2023\\\/06\\\/Daily-Ai_TL_colour.png\",\"width\":4501,\"height\":934,\"caption\":\"DailyAI\"},\"image\":{\"@id\":\"https:\\\/\\\/dailyai.com\\\/#\\\/schema\\\/logo\\\/image\\\/\"},\"sameAs\":[\"https:\\\/\\\/x.com\\\/DailyAIOfficial\",\"https:\\\/\\\/www.linkedin.com\\\/company\\\/dailyaiofficial\\\/\",\"https:\\\/\\\/www.youtube.com\\\/@DailyAIOfficial\"],\"legalName\":\"Zero Hour Investments FZCO\",\"foundingDate\":\"2023\",\"slogan\":\"The AI economy, measured.\",\"description\":\"DailyAI tracks the AI economy in numbers: data trackers, research and plain-English news on jobs, spending, pricing and funding.\",\"address\":{\"@type\":\"PostalAddress\",\"addressCountry\":\"AE\"},\"contactPoint\":{\"@type\":\"ContactPoint\",\"contactType\":\"customer support\",\"url\":\"https:\\\/\\\/dailyai.com\\\/contact\\\/\"},\"founder\":{\"@type\":\"Person\",\"name\":\"David Merry\",\"url\":\"https:\\\/\\\/dailyai.com\\\/author\\\/davidmerry\\\/\",\"sameAs\":[\"https:\\\/\\\/www.linkedin.com\\\/in\\\/davidmerry\\\/\"]}},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/#\\\/schema\\\/person\\\/285a97adb64b88751ac3440ece6b08a1\",\"name\":\"David Merry\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"fr-FR\",\"@id\":\"https:\\\/\\\/dailyai.com\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/david-merry-dailyai.jpg\",\"url\":\"https:\\\/\\\/dailyai.com\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/david-merry-dailyai.jpg\",\"contentUrl\":\"https:\\\/\\\/dailyai.com\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/david-merry-dailyai.jpg\",\"caption\":\"David Merry\"},\"description\":\"David Merry is the founder of DailyAI and a London-based entrepreneur. He co-founded Kyro, is a partner at Ki and writes about AI's impact on business, markets and everyday life.\",\"sameAs\":[\"https:\\\/\\\/www.linkedin.com\\\/in\\\/davidmerry\\\/\"],\"url\":\"https:\\\/\\\/dailyai.com\\\/fr\\\/author\\\/davidmerry\\\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Les agents de codage ne corrigent int\u00e9gralement que 26,4 % des modifications r\u00e9elles.","description":"Le test de co-conception d'Alibaba r\u00e9v\u00e8le que le meilleur agent de codage ne r\u00e9ussit que 26,4 % des 273 t\u00e2ches r\u00e9elles d'\u00e9dition d'applications, malgr\u00e9 le respect en moyenne de 69,9 % des crit\u00e8res.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/dailyai.com\/fr\/2026\/11\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\/","og_locale":"fr_FR","og_type":"article","og_title":"Coding Agents Fully Clear Only 26.4% of Real Edits","og_description":"Alibaba\u2019s Co-Design Bench finds the best coding agent fully clears only 26.4% of 273 real app-edit tasks, despite meeting 69.9% of criteria on average.","og_url":"https:\/\/dailyai.com\/fr\/2026\/11\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\/","og_site_name":"DailyAI","article_published_time":"2026-11-09T12:16:23+00:00","og_image":[{"width":1200,"height":630,"url":"https:\/\/dailyai.com\/wp-content\/uploads\/2026\/10\/dailyai-og-default.jpg","type":"image\/jpeg"}],"author":"David Merry","twitter_card":"summary_large_image","twitter_creator":"@DailyAIOfficial","twitter_site":"@DailyAIOfficial","twitter_misc":{"\u00c9crit par":"David Merry","Dur\u00e9e de lecture estim\u00e9e":"4 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"NewsArticle","@id":"https:\/\/dailyai.com\/2026\/11\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\/#article","isPartOf":{"@id":"https:\/\/dailyai.com\/2026\/11\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\/"},"author":{"name":"David Merry","@id":"https:\/\/dailyai.com\/#\/schema\/person\/285a97adb64b88751ac3440ece6b08a1"},"headline":"Best Coding Agent Fully Clears Only 26.4% of Real Tasks","datePublished":"2026-11-09T12:16:23+00:00","mainEntityOfPage":{"@id":"https:\/\/dailyai.com\/2026\/11\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\/"},"wordCount":906,"publisher":{"@id":"https:\/\/dailyai.com\/#organization"},"image":{"@id":"https:\/\/dailyai.com\/2026\/11\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\/#primaryimage"},"thumbnailUrl":"https:\/\/dailyai.com\/wp-content\/uploads\/2026\/11\/dai-house-15998.jpg","inLanguage":"fr-FR"},{"@type":"WebPage","@id":"https:\/\/dailyai.com\/2026\/11\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\/","url":"https:\/\/dailyai.com\/2026\/11\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\/","name":"Les agents de codage ne corrigent int\u00e9gralement que 26,4 % des modifications r\u00e9elles.","isPartOf":{"@id":"https:\/\/dailyai.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/dailyai.com\/2026\/11\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\/#primaryimage"},"image":{"@id":"https:\/\/dailyai.com\/2026\/11\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\/#primaryimage"},"thumbnailUrl":"https:\/\/dailyai.com\/wp-content\/uploads\/2026\/11\/dai-house-15998.jpg","datePublished":"2026-11-09T12:16:23+00:00","description":"Le test de co-conception d'Alibaba r\u00e9v\u00e8le que le meilleur agent de codage ne r\u00e9ussit que 26,4 % des 273 t\u00e2ches r\u00e9elles d'\u00e9dition d'applications, malgr\u00e9 le respect en moyenne de 69,9 % des crit\u00e8res.","breadcrumb":{"@id":"https:\/\/dailyai.com\/2026\/11\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\/#breadcrumb"},"inLanguage":"fr-FR","potentialAction":[{"@type":"ReadAction","target":["https:\/\/dailyai.com\/2026\/11\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\/"]}]},{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https:\/\/dailyai.com\/2026\/11\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\/#primaryimage","url":"https:\/\/dailyai.com\/wp-content\/uploads\/2026\/11\/dai-house-15998.jpg","contentUrl":"https:\/\/dailyai.com\/wp-content\/uploads\/2026\/11\/dai-house-15998.jpg","width":1472,"height":832,"caption":"Flat vector illustration of a cream clipboard with three navy checklist bars and one unfinished red bar beside a navy code window."},{"@type":"BreadcrumbList","@id":"https:\/\/dailyai.com\/2026\/11\/best-coding-agent-fully-clears-only-26-4-of-real-tasks\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/dailyai.com\/"},{"@type":"ListItem","position":2,"name":"Best Coding Agent Fully Clears Only 26.4% of Real Tasks"}]},{"@type":"WebSite","@id":"https:\/\/dailyai.com\/#website","url":"https:\/\/dailyai.com\/","name":"DailyAI","description":"L\u2019\u00e9conomie de l\u2019IA, mesur\u00e9e.","publisher":{"@id":"https:\/\/dailyai.com\/#organization"},"potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/dailyai.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"fr-FR"},{"@type":"Organization","@id":"https:\/\/dailyai.com\/#organization","name":"DailyAI","url":"https:\/\/dailyai.com\/","logo":{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https:\/\/dailyai.com\/#\/schema\/logo\/image\/","url":"https:\/\/dailyai.com\/wp-content\/uploads\/2023\/06\/Daily-Ai_TL_colour.png","contentUrl":"https:\/\/dailyai.com\/wp-content\/uploads\/2023\/06\/Daily-Ai_TL_colour.png","width":4501,"height":934,"caption":"DailyAI"},"image":{"@id":"https:\/\/dailyai.com\/#\/schema\/logo\/image\/"},"sameAs":["https:\/\/x.com\/DailyAIOfficial","https:\/\/www.linkedin.com\/company\/dailyaiofficial\/","https:\/\/www.youtube.com\/@DailyAIOfficial"],"legalName":"Zero Hour Investments FZCO","foundingDate":"2023","slogan":"The AI economy, measured.","description":"DailyAI suit l\u2019\u00e9conomie de l\u2019IA en chiffres : trackers de donn\u00e9es, \u00e9tudes et actualit\u00e9s en langage clair sur l\u2019emploi, les d\u00e9penses, les prix et les financements.","address":{"@type":"PostalAddress","addressCountry":"AE"},"contactPoint":{"@type":"ContactPoint","contactType":"customer support","url":"https:\/\/dailyai.com\/contact\/"},"founder":{"@type":"Person","name":"David Merry","url":"https:\/\/dailyai.com\/author\/davidmerry\/","sameAs":["https:\/\/www.linkedin.com\/in\/davidmerry\/"]}},{"@type":"Person","@id":"https:\/\/dailyai.com\/#\/schema\/person\/285a97adb64b88751ac3440ece6b08a1","name":"David Merry","image":{"@type":"ImageObject","inLanguage":"fr-FR","@id":"https:\/\/dailyai.com\/wp-content\/uploads\/2026\/10\/david-merry-dailyai.jpg","url":"https:\/\/dailyai.com\/wp-content\/uploads\/2026\/10\/david-merry-dailyai.jpg","contentUrl":"https:\/\/dailyai.com\/wp-content\/uploads\/2026\/10\/david-merry-dailyai.jpg","caption":"David Merry"},"description":"David Merry est le fondateur de DailyAI et un entrepreneur bas\u00e9 \u00e0 Londres. Il a cofond\u00e9 Kyro, est associ\u00e9 chez Ki et \u00e9crit sur l'impact de l'IA sur les entreprises, les march\u00e9s et la vie quotidienne.","sameAs":["https:\/\/www.linkedin.com\/in\/davidmerry\/"],"url":"https:\/\/dailyai.com\/fr\/author\/davidmerry\/"}]}},"_links":{"self":[{"href":"https:\/\/dailyai.com\/fr\/wp-json\/wp\/v2\/posts\/15998","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/dailyai.com\/fr\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/dailyai.com\/fr\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/dailyai.com\/fr\/wp-json\/wp\/v2\/users\/16"}],"replies":[{"embeddable":true,"href":"https:\/\/dailyai.com\/fr\/wp-json\/wp\/v2\/comments?post=15998"}],"version-history":[{"count":1,"href":"https:\/\/dailyai.com\/fr\/wp-json\/wp\/v2\/posts\/15998\/revisions"}],"predecessor-version":[{"id":16000,"href":"https:\/\/dailyai.com\/fr\/wp-json\/wp\/v2\/posts\/15998\/revisions\/16000"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/dailyai.com\/fr\/wp-json\/wp\/v2\/media\/15999"}],"wp:attachment":[{"href":"https:\/\/dailyai.com\/fr\/wp-json\/wp\/v2\/media?parent=15998"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/dailyai.com\/fr\/wp-json\/wp\/v2\/categories?post=15998"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/dailyai.com\/fr\/wp-json\/wp\/v2\/tags?post=15998"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}