[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-post-\u002Fnews\u002Fopenai-pando-git-backend-monorepo-codex-agents":3,"news-related-en-\u002Fnews\u002Fopenai-pando-git-backend-monorepo-codex-agents":225,"hreflang-openai-pando-git-backend-2026-11":246},{"doc":4,"redirect":8},{"id":5,"title":6,"author":7,"author_avatar":8,"body":9,"category":203,"date":204,"description":205,"extension":206,"image":207,"meta":208,"navigation":209,"path":210,"seo":211,"seoDescription":212,"seoTitle":213,"stem":214,"tags":215,"translationKey":223,"updatedAt":8,"__hash__":224},"news_en\u002Fnews\u002Fopenai-pando-git-backend-monorepo-codex-agents.md","OpenAI rebuilt the Git backend behind its monorepo to keep up with Codex agents","Paul Lefizelier",null,{"type":10,"value":11,"toc":193},"minimark",[12,25,34,39,42,56,59,63,66,73,85,88,91,95,103,119,122,126,129,132,136,157,160,164,167,170,190],[13,14,15,16,24],"p",{},"OpenAI has published a detailed account of how it replaced the system serving its main code repository. In ",[17,18,23],"a",{"href":19,"rel":20,"target":22},"https:\u002F\u002Fopenai.com\u002Findex\u002Fscaling-git-at-openai\u002F",[21],"nofollow","_blank","Scaling Git at OpenAI",", dated November 2, Aaron Friel, Taylor Blau and Ted Nyman describe Pando, a new backend for Git repositories that \"now powers all development in OpenAI's monorepo\". The reason they give is blunt: the volume of parallel work that Codex makes possible had outgrown the previous setup.",[13,26,27,28,33],{},"The post reached the ",[17,29,32],{"href":30,"rel":31,"target":22},"https:\u002F\u002Fnews.ycombinator.com\u002Fitem?id=50345924",[21],"Hacker News front page"," within hours. It matters beyond OpenAI because it is one of the first detailed, first-hand descriptions of what coding agents do to source control once they are used at scale.",[35,36,38],"h2",{"id":37},"what-broke","What broke",[13,40,41],{},"According to OpenAI, Codex lets each engineer run far more work in parallel, which means more branches, more pushes, more merges and more fetches from development environments, research jobs and CI. \"One engineer now has several changes ready for CI at once, and each push can set several jobs fetching the same commit,\" the authors write.",[13,43,44,45,49,50,55],{},"Starting around the end of 2025, the systems behind the monorepo \"were struggling to keep up\", with slow fetches and pushes, failed merges, and \"reliability at times falling below 90%\", ",[17,46,48],{"href":19,"rel":47,"target":22},[21],"the post says",". ",[17,51,54],{"href":52,"rel":53,"target":22},"https:\u002F\u002Fthenewstack.io\u002Fopenai-pando-git-backend\u002F",[21],"The New Stack reports"," that, according to OpenAI, quarterly commits to the monorepo will have grown by more than 2,400% by the end of this year compared with the first quarter of 2024.",[13,57,58],{},"The team's question was whether it had hit a limit of Git itself or of the way it served the repository. Its answer was the second, so it kept the Git protocol and client and rebuilt the server side.",[35,60,62],{"id":61},"how-pando-works","How Pando works",[13,64,65],{},"OpenAI says Pando drops two things most Git servers treat as requirements.",[13,67,68,72],{},[69,70,71],"strong",{},"No repository-wide snapshot."," Conventional designs order all reference updates across the whole repository, so a push to one branch can wait on a push to an unrelated one. Pando coordinates each branch independently. The only contended key is the branch pointer, protected by an atomic compare-and-swap style operation. OpenAI calls the result causal consistency: updates to a single branch are serialized, but a client reading several branches may see them at slightly different points in time. \"In practice, users and agents are agnostic to the global ordering of ref updates,\" the authors argue.",[13,74,75,78,79,84],{},[69,76,77],{},"No custom packfile per client."," A normal Git server computes, for each fetch, the objects the client lacks and compresses a fresh packfile. Pando instead stores every pushed packfile as an immutable blob, records each new branch tip and its dependencies in a metadata store, and on fetch selects stored packs that cover what the client needs. It hands back download links through ",[17,80,83],{"href":81,"rel":82,"target":22},"https:\u002F\u002Fgit-scm.com\u002Fdocs\u002Fpackfile-uri",[21],"packfile-uri",", an existing but rarely used Git protocol feature, and a cache service called Express serves the bytes from regional caches with hit rates above 99%.",[13,86,87],{},"To make packs reusable, Pando advertises only the trunk branch when a client pushes a new feature branch, rather than listing every ref. OpenAI notes that with millions of branches the full ref advertisement alone can exceed 100 MiB. The tradeoff is that repeat pushes to a branch resend everything not on trunk; in OpenAI's monorepo the median pack produced this way was under 20 KiB and the 99th percentile under 3 MiB.",[13,89,90],{},"Because thousands of small packs per day would slow clients down, background workers merge them into larger \"covers\", in a way the post compares to compaction in log-structured merge-trees. The monorepo currently uses up to four levels of covers, which also let Pando plan a fetch in logarithmic time relative to how far behind a client is.",[35,92,94],{"id":93},"the-numbers-openai-reports","The numbers OpenAI reports",[13,96,97,98,102],{},"All figures below come from ",[17,99,101],{"href":19,"rel":100,"target":22},[21],"OpenAI's post"," and have not been independently measured:",[104,105,106,110,113,116],"ul",{},[107,108,109],"li",{},"median push latency cut by nearly 5x (the post also phrases it as nearly 80% lower);",[107,111,112],{},"close to a petabyte of Git data delivered per day, \"with just a handful of CPU cores\";",[107,114,115],{},"more than 57 terabytes of packs served in a single hour, over 99% of it from local disk cache;",[107,117,118],{},"clones and fetches saturating 10 Gbps links on standard cloud virtual machines.",[13,120,121],{},"The interactive simulation in the post comparing Pando with three-phase commit and write-ahead-log designs is labelled by OpenAI as \"synthetic teaching models, not production benchmarks\".",[35,123,125],{"id":124},"built-with-agents-in-10-weeks","Built with agents, in 10 weeks",[13,127,128],{},"OpenAI says it went from kickoff to production in 10 weeks, with the core \"designed by a single engineer, implemented by teams of agents, and deployed by a handful of engineers\". Pando is written in Rust and has served production traffic since April. The team wrote a Git conformance test suite so agents could check their changes against standard Git behavior, and used deterministic simulation testing to find race conditions. It also says agents running GPT-6 Astra wrote Lean and TLA+ specifications to formalize a \"theory of Pando\", which were used to find bugs and prove properties of the system.",[13,130,131],{},"Migration was gradual. Pando first ran as a read- and write-through proxy in front of the old system, and OpenAI now ships its own Git build to developer machines and inside the ChatGPT desktop app. Developers could ask Codex to migrate their own environments using dedicated skills.",[35,133,135],{"id":134},"what-goes-back-to-git-and-what-does-not","What goes back to Git, and what does not",[13,137,138,139,144,145,150,151,156],{},"OpenAI says it has sent client fixes back to the Git project, covering concurrent packfile-uri downloads, cleanup after interrupted transfers and new tracing, and that several are already in upstream Git. Ted Nyman's ",[17,140,143],{"href":141,"rel":142,"target":22},"https:\u002F\u002Fpublic-inbox.org\u002Fgit\u002Fcover.1784874850.git.tnyman@openai.com\u002F",[21],"patch series on concurrent downloads"," is public on the Git mailing list. Other changes, such as parallel pack indexing and \"explicit haves\", live for now in OpenAI's ",[17,146,149],{"href":147,"rel":148,"target":22},"https:\u002F\u002Fgithub.com\u002Fopenai\u002Fgit",[21],"public Git fork",", tracked on a ",[17,152,155],{"href":153,"rel":154,"target":22},"https:\u002F\u002Foai-git.openai.chatgpt.site\u002F",[21],"patch status page"," that, when we checked, listed 50 patches for upstream and 92 commits carried on top of upstream Git.",[13,158,159],{},"Pando itself is not open source. OpenAI only says it is \"exploring ways to share the technology that makes Pando possible with our users\". The New Stack notes that Anthropic has also built its own source-control system, Scion, which it has no plans to release, and that startups such as Entire and Cursor's Origin project are working on the same problem.",[35,161,163],{"id":162},"why-it-matters-for-teams-running-coding-agents","Why it matters for teams running coding agents",[13,165,166],{},"Few teams have OpenAI's volume, but the failure pattern it describes scales down. When each developer runs several agents, the bottleneck moves from writing code to everything around it: pushes landing at the same time, CI jobs cloning the same commit, branch counts growing fast. OpenAI's diagnosis is that the Git client and protocol held up and the server design did not.",[13,168,169],{},"Some practical takeaways from the post:",[104,171,172,178,184],{},[107,173,174,177],{},[69,175,176],{},"Watch contention, not just storage."," Systems that order every write across a repository make unrelated agent branches wait on each other.",[107,179,180,183],{},[69,181,182],{},"Reads multiply faster than writes."," One push can trigger several fetches; caching identical responses mattered more for OpenAI than shrinking each pack.",[107,185,186,189],{},[69,187,188],{},"Staying on Git paid off."," Keeping the protocol meant no retraining for engineers or agents, and every client fix can benefit the wider ecosystem.",[13,191,192],{},"The open question is whether GitHub and other hosts will adopt similar tradeoffs, such as relaxing repository-wide consistency, before agent-heavy teams start building their own backends.",{"title":194,"searchDepth":195,"depth":195,"links":196},"",2,[197,198,199,200,201,202],{"id":37,"depth":195,"text":38},{"id":61,"depth":195,"text":62},{"id":93,"depth":195,"text":94},{"id":124,"depth":195,"text":125},{"id":134,"depth":195,"text":135},{"id":162,"depth":195,"text":163},"tools","2026-11-03T10:15:00","OpenAI has detailed Pando, a Git-compatible backend it built in 10 weeks with Codex. It coordinates each branch separately, reuses packfiles across fetches, cut median push latency by nearly 5x and now serves close to a petabyte of Git data a day.","md","\u002Fimages\u002Fnews\u002Fopenai-pando-git-backend-monorepo-codex-agents.jpg",{},true,"\u002Fnews\u002Fopenai-pando-git-backend-monorepo-codex-agents",{"title":6,"description":205},"OpenAI's Pando keeps Git but coordinates each branch separately and reuses packfiles, cutting median push latency nearly 5x. What it means for agent-heavy teams.","OpenAI Pando: a new Git backend built for Codex-scale pushes","news\u002Fopenai-pando-git-backend-monorepo-codex-agents",[216,217,218,219,220,221,222],"openai","git","codex","source-control","monorepo","ai-coding-agents","infrastructure","openai-pando-git-backend-2026-11","i62y9TsvCEVKl7TfzXkp83_DSRtx7orK4kCJ0dkomeU",[226,233,239],{"title":227,"description":228,"path":229,"date":230,"image":231,"category":232},"Anthropic's IPO filing says it may soon cross its own threshold for AI-accelerated AI research","In its amended S-1, Anthropic added a risk factor saying internal estimates for future models indicate it may \"imminently\" meet its Responsible Scaling Policy's definition of \"dramatically accelerated progress attributable to AI\". No current model has crossed it. Here is what the threshold means and what the filing says about Claude building Claude.","\u002Fnews\u002Fanthropic-s1-recursive-self-improvement-rsp-threshold-ai-rd","2026-11-03T10:20:00","\u002Fimages\u002Fnews\u002Fanthropic-s1-recursive-self-improvement-rsp-threshold-ai-rd.jpg","ai",{"title":234,"description":235,"path":236,"date":237,"image":238,"category":232},"MIT researchers show GPT-6 Astra can read and edit neural network weights directly","In a blog post published November 2, an MIT CSAIL team reports that GPT-6 Astra can identify what small networks were trained to do from their raw weights, write working game-playing networks from scratch, and lift Qwen3.5-0.8B from 7.9% to 48.1% on MATH-500 by editing 0.016% of its weights, without gradient descent.","\u002Fnews\u002Fmit-weight-space-programming-gpt-6-astra-edits-neural-network-weights","2026-11-03T08:10:00","\u002Fimages\u002Fnews\u002Fmit-weight-space-programming-gpt-6-astra-edits-neural-network-weights.jpg",{"title":240,"description":241,"path":242,"date":243,"image":244,"category":245},"DeepSeek open-sources dsh-superagent, a coding agent that splits routing from execution","DeepSeek's Harness team has published dsh-superagent, an MIT-licensed research prototype that turns one GitHub repository into a proactive coding agent: a long-running Router takes requests and hands each one to an isolated Worker session. Here is how it works, what DeepSeek measured and what to watch before running it.","\u002Fnews\u002Fdeepseek-dsh-superagent-open-source-router-worker-coding-agent","2026-11-02T20:20:00+01:00","\u002Fdeepseek-dsh-superagent-open-source-router-worker-coding-agent.jpg","vibecoding",{"en":247,"fr":366,"es":8},{"id":5,"title":6,"author":7,"author_avatar":8,"body":248,"category":203,"date":204,"description":205,"extension":206,"image":207,"meta":363,"navigation":209,"path":210,"seo":364,"seoDescription":212,"seoTitle":213,"stem":214,"tags":365,"translationKey":223,"updatedAt":8,"__hash__":224},{"type":10,"value":249,"toc":355},[250,255,260,262,264,272,274,276,278,282,289,291,293,295,300,310,312,314,316,318,320,331,333,335,337,339,353],[13,251,15,252,24],{},[17,253,23],{"href":19,"rel":254,"target":22},[21],[13,256,27,257,33],{},[17,258,32],{"href":30,"rel":259,"target":22},[21],[35,261,38],{"id":37},[13,263,41],{},[13,265,44,266,49,269,55],{},[17,267,48],{"href":19,"rel":268,"target":22},[21],[17,270,54],{"href":52,"rel":271,"target":22},[21],[13,273,58],{},[35,275,62],{"id":61},[13,277,65],{},[13,279,280,72],{},[69,281,71],{},[13,283,284,78,286,84],{},[69,285,77],{},[17,287,83],{"href":81,"rel":288,"target":22},[21],[13,290,87],{},[13,292,90],{},[35,294,94],{"id":93},[13,296,97,297,102],{},[17,298,101],{"href":19,"rel":299,"target":22},[21],[104,301,302,304,306,308],{},[107,303,109],{},[107,305,112],{},[107,307,115],{},[107,309,118],{},[13,311,121],{},[35,313,125],{"id":124},[13,315,128],{},[13,317,131],{},[35,319,135],{"id":134},[13,321,138,322,144,325,150,328,156],{},[17,323,143],{"href":141,"rel":324,"target":22},[21],[17,326,149],{"href":147,"rel":327,"target":22},[21],[17,329,155],{"href":153,"rel":330,"target":22},[21],[13,332,159],{},[35,334,163],{"id":162},[13,336,166],{},[13,338,169],{},[104,340,341,345,349],{},[107,342,343,177],{},[69,344,176],{},[107,346,347,183],{},[69,348,182],{},[107,350,351,189],{},[69,352,188],{},[13,354,192],{},{"title":194,"searchDepth":195,"depth":195,"links":356},[357,358,359,360,361,362],{"id":37,"depth":195,"text":38},{"id":61,"depth":195,"text":62},{"id":93,"depth":195,"text":94},{"id":124,"depth":195,"text":125},{"id":134,"depth":195,"text":135},{"id":162,"depth":195,"text":163},{},{"title":6,"description":205},[216,217,218,219,220,221,222],{"id":367,"title":368,"author":7,"author_avatar":8,"body":369,"category":203,"date":204,"description":542,"extension":206,"image":207,"meta":543,"navigation":209,"path":544,"seo":545,"seoDescription":546,"seoTitle":547,"stem":548,"tags":549,"translationKey":223,"updatedAt":8,"__hash__":552},"news_fr\u002Ffr\u002Fnews\u002Fopenai-pando-backend-git-monorepo-agents-codex.md","OpenAI a reconstruit le backend Git de son monorepo pour suivre le rythme des agents Codex",{"type":10,"value":370,"toc":534},[371,378,386,390,393,405,408,412,415,421,431,434,437,441,449,463,466,470,473,476,480,498,501,505,508,511,531],[13,372,373,374,377],{},"OpenAI a publié un récit détaillé du remplacement du système qui sert son principal dépôt de code. Dans ",[17,375,23],{"href":19,"rel":376,"target":22},[21],", daté du 2 novembre, Aaron Friel, Taylor Blau et Ted Nyman présentent Pando, un nouveau backend pour dépôts Git qui « alimente désormais tout le développement dans le monorepo d'OpenAI ». La raison invoquée est simple : le volume de travail en parallèle que permet Codex avait dépassé ce que l'ancienne infrastructure pouvait encaisser.",[13,379,380,381,385],{},"Le billet est arrivé en quelques heures en ",[17,382,384],{"href":30,"rel":383,"target":22},[21],"première page de Hacker News",". Il intéresse bien au-delà d'OpenAI, car c'est l'une des premières descriptions détaillées, de première main, de ce que les agents de code font subir à la gestion de versions quand on les utilise à grande échelle.",[35,387,389],{"id":388},"ce-qui-a-cassé","Ce qui a cassé",[13,391,392],{},"Selon OpenAI, Codex permet à chaque ingénieur de mener beaucoup plus de travail en parallèle : plus de branches, plus de push, plus de merges et plus de fetch depuis les environnements de développement, les jobs de recherche et la CI. « Un ingénieur a désormais plusieurs changements prêts pour la CI en même temps, et chaque push peut déclencher plusieurs jobs qui récupèrent le même commit », écrivent les auteurs.",[13,394,395,396,49,400,404],{},"À partir de la fin 2025 environ, les systèmes derrière le monorepo « peinaient à suivre » : fetch et push lents, merges en échec, et une fiabilité « tombant parfois sous les 90 % », ",[17,397,399],{"href":19,"rel":398,"target":22},[21],"indique le billet",[17,401,403],{"href":52,"rel":402,"target":22},[21],"The New Stack rapporte"," que, d'après OpenAI, le nombre de commits trimestriels dans le monorepo aura progressé de plus de 2 400 % d'ici la fin de l'année par rapport au premier trimestre 2024.",[13,406,407],{},"L'équipe s'est demandé si elle avait atteint une limite de Git lui-même ou de la façon dont elle servait le dépôt. Sa réponse : la seconde. Elle a donc conservé le protocole et le client Git et reconstruit le côté serveur.",[35,409,411],{"id":410},"comment-fonctionne-pando","Comment fonctionne Pando",[13,413,414],{},"OpenAI explique que Pando renonce à deux choses que la plupart des serveurs Git considèrent comme indispensables.",[13,416,417,420],{},[69,418,419],{},"Pas d'instantané global du dépôt."," Les architectures classiques ordonnent toutes les mises à jour de références sur l'ensemble du dépôt, si bien qu'un push sur une branche peut attendre un push sur une branche sans rapport. Pando coordonne chaque branche indépendamment. La seule clé disputée est le pointeur de branche, protégé par une opération atomique de type compare-and-swap. OpenAI parle de cohérence causale : les mises à jour d'une même branche sont sérialisées, mais un client qui lit plusieurs branches peut les voir à des instants légèrement différents. « En pratique, les utilisateurs et les agents ne dépendent pas de l'ordre global des mises à jour de références », défendent les auteurs.",[13,422,423,426,427,430],{},[69,424,425],{},"Pas de packfile sur mesure pour chaque client."," Un serveur Git classique calcule, à chaque fetch, les objets qui manquent au client et compresse un nouveau packfile. Pando stocke au contraire chaque packfile poussé comme un blob immuable, enregistre chaque nouvelle tête de branche et ses dépendances dans une base de métadonnées, puis, au fetch, sélectionne les packs déjà stockés qui couvrent les besoins du client. Il renvoie des liens de téléchargement via ",[17,428,83],{"href":81,"rel":429,"target":22},[21],", une fonction du protocole Git qui existe depuis longtemps mais reste peu utilisée, et un service de cache baptisé Express sert les octets depuis des caches régionaux, avec un taux de succès supérieur à 99 %.",[13,432,433],{},"Pour rendre les packs réutilisables, Pando n'annonce que la branche principale (trunk) quand un client pousse une nouvelle branche de fonctionnalité, au lieu de lister toutes les références. OpenAI précise qu'avec des millions de branches, la seule annonce complète des références peut dépasser 100 Mio. Contrepartie : les push répétés sur une même branche renvoient tout ce qui n'est pas sur trunk. Dans le monorepo d'OpenAI, le pack médian produit ainsi faisait moins de 20 Kio, et le 99e centile moins de 3 Mio.",[13,435,436],{},"Comme des milliers de petits packs par jour ralentiraient les clients, des workers en arrière-plan les fusionnent en « couvertures » plus grandes, un mécanisme que le billet compare au compactage des arbres LSM (log-structured merge-trees). Le monorepo utilise aujourd'hui jusqu'à quatre niveaux de couvertures, ce qui permet aussi à Pando de planifier un fetch en temps logarithmique selon le retard du client.",[35,438,440],{"id":439},"les-chiffres-annoncés-par-openai","Les chiffres annoncés par OpenAI",[13,442,443,444,448],{},"Tous les chiffres ci-dessous viennent du ",[17,445,447],{"href":19,"rel":446,"target":22},[21],"billet d'OpenAI"," et n'ont pas été mesurés de façon indépendante :",[104,450,451,454,457,460],{},[107,452,453],{},"latence médiane des push divisée par près de 5 (le billet parle aussi d'une baisse de près de 80 %) ;",[107,455,456],{},"près d'un pétaoctet de données Git servi par jour, « avec une poignée de cœurs CPU » ;",[107,458,459],{},"plus de 57 téraoctets de packs servis en une seule heure, dont plus de 99 % depuis le cache disque local ;",[107,461,462],{},"des clones et fetch qui saturent des liens à 10 Gbit\u002Fs sur des machines virtuelles cloud standard.",[13,464,465],{},"La simulation interactive du billet, qui compare Pando à des architectures à validation en trois phases et à journal d'écriture anticipée, est présentée par OpenAI comme des « modèles pédagogiques synthétiques, pas des benchmarks de production ».",[35,467,469],{"id":468},"construit-avec-des-agents-en-10-semaines","Construit avec des agents, en 10 semaines",[13,471,472],{},"OpenAI affirme être passé du lancement à la production en 10 semaines, avec un cœur « conçu par un seul ingénieur, implémenté par des équipes d'agents et déployé par une poignée d'ingénieurs ». Pando est écrit en Rust et sert du trafic de production depuis avril. L'équipe a écrit une suite de tests de conformité Git pour que les agents puissent vérifier leurs modifications par rapport au comportement standard de Git, et a utilisé des tests par simulation déterministe pour débusquer des conditions de concurrence. Elle indique aussi que des agents sous GPT-6 Astra ont écrit des spécifications en Lean et TLA+ pour formaliser une « théorie de Pando », utilisée pour trouver des bugs et prouver des propriétés du système.",[13,474,475],{},"La migration s'est faite progressivement. Pando a d'abord tourné comme proxy en lecture et en écriture devant l'ancien système, et OpenAI distribue désormais sa propre version de Git sur les postes des développeurs et dans l'application ChatGPT pour ordinateur. Les développeurs pouvaient demander à Codex de migrer leurs environnements grâce à des skills dédiés.",[35,477,479],{"id":478},"ce-qui-revient-à-git-et-ce-qui-reste-chez-openai","Ce qui revient à Git, et ce qui reste chez OpenAI",[13,481,482,483,487,488,492,493,497],{},"OpenAI dit avoir renvoyé au projet Git des correctifs côté client, sur les téléchargements packfile-uri concurrents, le nettoyage après un transfert interrompu et un nouveau traçage, et que plusieurs sont déjà intégrés à Git. La ",[17,484,486],{"href":141,"rel":485,"target":22},[21],"série de patchs de Ted Nyman sur les téléchargements concurrents"," est publique sur la liste de diffusion de Git. D'autres changements, comme l'indexation des packs en parallèle et les « explicit haves », vivent pour l'instant dans le ",[17,489,491],{"href":147,"rel":490,"target":22},[21],"fork public de Git d'OpenAI",", suivi sur une ",[17,494,496],{"href":153,"rel":495,"target":22},[21],"page d'état des patchs"," qui, lors de notre vérification, recensait 50 patchs destinés à Git et 92 commits appliqués par-dessus la version officielle.",[13,499,500],{},"Pando lui-même n'est pas open source. OpenAI dit seulement « étudier des moyens de partager la technologie qui rend Pando possible avec nos utilisateurs ». The New Stack note qu'Anthropic a aussi construit son propre système de gestion de versions, Scion, sans projet de le publier, et que des startups comme Entire ou le projet Origin de Cursor s'attaquent au même problème.",[35,502,504],{"id":503},"ce-que-ça-change-pour-les-équipes-qui-utilisent-des-agents-de-code","Ce que ça change pour les équipes qui utilisent des agents de code",[13,506,507],{},"Peu d'équipes ont le volume d'OpenAI, mais le schéma de panne décrit se retrouve à plus petite échelle. Quand chaque développeur fait tourner plusieurs agents, le goulot d'étranglement se déplace de l'écriture du code vers tout ce qui l'entoure : push simultanés, jobs de CI qui clonent le même commit, nombre de branches qui explose. Le diagnostic d'OpenAI : le client et le protocole Git ont tenu, l'architecture du serveur non.",[13,509,510],{},"Quelques enseignements concrets du billet :",[104,512,513,519,525],{},[107,514,515,518],{},[69,516,517],{},"Surveiller la contention, pas seulement le stockage."," Les systèmes qui ordonnent chaque écriture sur tout le dépôt font attendre des branches d'agents qui n'ont rien à voir entre elles.",[107,520,521,524],{},[69,522,523],{},"Les lectures se multiplient plus vite que les écritures."," Un push peut déclencher plusieurs fetch ; pour OpenAI, mettre en cache des réponses identiques a compté davantage que réduire la taille de chaque pack.",[107,526,527,530],{},[69,528,529],{},"Rester sur Git a payé."," Garder le protocole évitait de former à nouveau ingénieurs et agents, et chaque correctif côté client peut profiter à tout l'écosystème.",[13,532,533],{},"Reste à savoir si GitHub et les autres hébergeurs adopteront des compromis similaires, comme l'abandon de la cohérence globale du dépôt, avant que les équipes les plus équipées en agents ne se mettent à construire leurs propres backends.",{"title":194,"searchDepth":195,"depth":195,"links":535},[536,537,538,539,540,541],{"id":388,"depth":195,"text":389},{"id":410,"depth":195,"text":411},{"id":439,"depth":195,"text":440},{"id":468,"depth":195,"text":469},{"id":478,"depth":195,"text":479},{"id":503,"depth":195,"text":504},"OpenAI détaille Pando, un backend compatible Git développé en 10 semaines avec Codex. Il coordonne chaque branche séparément, réutilise les packfiles d'un fetch à l'autre, a divisé par près de 5 la latence médiane des push et sert près d'un pétaoctet de données Git par jour.",{},"\u002Ffr\u002Fnews\u002Fopenai-pando-backend-git-monorepo-agents-codex",{"title":368,"description":542},"Pando, le backend d'OpenAI, garde le protocole Git mais coordonne chaque branche séparément et réutilise les packfiles. Latence médiane des push divisée par près de 5.","OpenAI Pando : un nouveau backend Git pensé pour les agents Codex","fr\u002Fnews\u002Fopenai-pando-backend-git-monorepo-agents-codex",[216,217,218,550,220,551,222],"gestion-de-versions","agents-de-code","DjqhPWSKR12v43C5JmfIMquShR2avEyvp72ZO1Ygvq4"]