[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-post-\u002Ffr\u002Fnews\u002Fdeepseek-dsh-superagent-agent-de-code-open-source-router-worker":3,"news-related-fr-\u002Ffr\u002Fnews\u002Fdeepseek-dsh-superagent-agent-de-code-open-source-router-worker":323,"hreflang-deepseek-dsh-superagent-open-source-router-worker-coding-agent":344},{"id":4,"title":5,"author":6,"author_avatar":7,"body":8,"category":302,"date":303,"description":304,"extension":305,"image":306,"meta":307,"navigation":308,"path":309,"seo":310,"seoDescription":311,"seoTitle":312,"stem":313,"tags":314,"translationKey":321,"updatedAt":303,"__hash__":322},"news_fr\u002Ffr\u002Fnews\u002Fdeepseek-dsh-superagent-agent-de-code-open-source-router-worker.md","DeepSeek publie en open source dsh-superagent, un agent de code qui sépare le routage de l'exécution","Paul Lefizelier",null,{"type":9,"value":10,"toc":292},"minimark",[11,46,53,58,66,69,73,87,103,112,121,125,134,186,195,199,210,213,217,229,232,246,250,258,261,286],[12,13,14,15,19,20,28,29,33,34,39,40,45],"p",{},"Le ",[16,17,18],"strong",{},"31 octobre 2026",", l'équipe DeepSeek Harness a publié ",[21,22,27],"a",{"href":23,"rel":24,"target":26},"https:\u002F\u002Fgithub.com\u002Fdeepseek-ai\u002Fdsh-superagent-prototype",[25],"nofollow","_blank","dsh-superagent-prototype"," sur GitHub sous licence MIT, accompagné d'un article de conception intitulé ",[30,31,32],"em",{},"Exploring Super-agent Design: Towards a More Autonomous Agent System"," (",[21,35,38],{"href":36,"rel":37,"target":26},"https:\u002F\u002Foffloop.org\u002Flearn\u002Fexploring-deepseek-super-agent-design",[25],"traduction anglaise","). Lundi soir, le dépôt dépassait les 12 000 étoiles. ",[21,41,44],{"href":42,"rel":43,"target":26},"https:\u002F\u002Fwww.marktechpost.com\u002F2026\u002F11\u002F01\u002Fdeepseek-dsh-superagent-router-worker-coding-agent\u002F",[25],"MarkTechPost"," le décrit comme un bundle TypeScript en version 0.1.0, publié en un seul commit.",[12,47,14,48,52],{},[21,49,51],{"href":23,"rel":50,"target":26},[25],"README"," est clair sur son statut : il s'agit d'un « prototype de recherche fourni à titre de référence, pas d'un produit officiel de DeepSeek », écrit en grande partie avec des agents de code, et différent à la fois de la version interne de DeepSeek et du système décrit dans l'article. Il mérite pourtant le détour pour qui construit avec des agents de code, car il s'attaque à un problème que la plupart des outils actuels laissent à l'utilisateur : gérer plusieurs sessions d'agent en même temps.",[54,55,57],"h2",{"id":56},"le-problème-les-sessions-ne-passent-pas-à-léchelle-humaine","Le problème : les sessions ne passent pas à l'échelle humaine",[12,59,60,61,65],{},"La plupart des agents de code considèrent la session comme l'unité de travail. Dans l'article de conception, l'équipe de DeepSeek écrit qu'« une personne ne peut généralement suivre que trois à cinq sessions actives à la fois ». La limite du travail en parallèle devient donc l'attention humaine plutôt que la capacité du modèle (",[21,62,64],{"href":36,"rel":63,"target":26},[25],"traduction",").",[12,67,68],{},"Leur réponse est ce qu'ils appellent un Super-agent : un agent unique et continu, que l'on sollicite là où l'on travaille déjà, dans un commentaire GitHub ou une messagerie, pendant que le système décide en coulisses comment répartir le travail. L'article lui fixe quatre propriétés : continuité, efficacité, proactivité et persistance.",[54,70,72],{"id":71},"comment-fonctionne-le-prototype","Comment fonctionne le prototype",[12,74,75,76,81,82,86],{},"dsh-superagent est distribué comme un bundle pour ",[21,77,80],{"href":78,"rel":79,"target":26},"https:\u002F\u002Fwww.deepseek.com\u002Fen\u002Fharness\u002F",[25],"DeepSeek Harness"," et surveille un dépôt GitHub, avec Feishu comme source de messagerie optionnelle. ",[21,83,85],{"href":42,"rel":84,"target":26},[25],"L'architecture compte deux couches"," :",[88,89,90,97],"ul",{},[91,92,93,96],"li",{},[16,94,95],{},"Un Router"," (la session principale) tourne en continu, lit les événements entrants et décide s'ils demandent un nouveau travail. Il utilise un preset restreint, avec moins d'outils.",[91,98,99,102],{},[16,100,101],{},"Des Workers"," (un par tâche, appelée Ticket) disposent chacun de leur propre session d'agent et de leur propre répertoire, écrivent le code et répondent dans la conversation d'origine, par exemple avec une pull request.",[12,104,105,106,111],{},"Chaque Ticket porte un Mandate immuable qui consigne pourquoi le travail peut avoir lieu. Les commentaires ultérieurs sur une pull request parviennent au bon Worker grâce à une Subscription, sans qu'il soit nécessaire de mentionner l'agent à nouveau. Selon la ",[21,107,110],{"href":108,"rel":109,"target":26},"https:\u002F\u002Fgithub.com\u002Fdeepseek-ai\u002Fdsh-superagent-prototype\u002Fblob\u002Fmain\u002Fdocs\u002F003-execution.md",[25],"documentation d'exécution",", les commandes shell au premier plan passent en arrière-plan au bout de cinq secondes, si bien qu'un Worker occupé par un long build peut continuer à recevoir des messages.",[12,113,114,115,120],{},"Un panneau web local n'affiche que ce qui demande une décision humaine. C'est l'application de la règle posée dans le ",[21,116,119],{"href":117,"rel":118,"target":26},"https:\u002F\u002Fgithub.com\u002Fdeepseek-ai\u002Fdsh-superagent-prototype\u002Fblob\u002Fmain\u002Fdocs\u002F000-philosophy.md",[25],"document de philosophie"," : « Prendre le moins d'attention humaine possible, et apporter le plus de valeur possible. » L'agent confirme une demande une fois, reste silencieux pendant qu'il travaille, et ne réécrit qu'avec un résultat, une question ou un échec.",[54,122,124],{"id":123},"le-travail-proactif-désactivé-de-fait-par-défaut","Le travail proactif, désactivé de fait par défaut",[12,126,127,128,133],{},"L'agent peut prendre en charge des tâches que personne n'a explicitement demandées, mais uniquement selon des règles écrites par l'équipe. Le ",[21,129,132],{"href":130,"rel":131,"target":26},"https:\u002F\u002Fgithub.com\u002Fdeepseek-ai\u002Fdsh-superagent-prototype\u002Fblob\u002Fmain\u002Fdocs\u002F002-authorization.md",[25],"document d'autorisation"," définit trois modes d'acceptation :",[135,136,137,150],"table",{},[138,139,140],"thead",{},[141,142,143,147],"tr",{},[144,145,146],"th",{},"Mode",[144,148,149],{},"Effet",[151,152,153,165,176],"tbody",{},[141,154,155,162],{},[156,157,158],"td",{},[159,160,161],"code",{},"off",[156,163,164],{},"Aucune demande déduite ne devient candidate",[141,166,167,173],{},[156,168,169,172],{},[159,170,171],{},"observe"," (par défaut)",[156,174,175],{},"Les candidates déduites sont évaluées et les avis positifs consignés comme « aurait accepté », mais rien n'est accepté",[141,177,178,183],{},[156,179,180],{},[159,181,182],{},"auto",[156,184,185],{},"Une évaluation positive déclenche le travail",[12,187,188,189,191,192,194],{},"Les demandes explicites des personnes autorisées fonctionnent dans tous les modes. La documentation recommande de démarrer en ",[159,190,171],{}," et de ne passer en ",[159,193,182],{}," qu'après avoir relu ce que l'agent aurait fait. Le même document précise que les messages, commentaires et documents cités « ne peuvent pas modifier » les permissions appliquées par l'hôte, ce qui compte face aux injections de prompt via les commentaires d'issues.",[54,196,198],{"id":197},"ce-que-deepseek-a-mesuré","Ce que DeepSeek a mesuré",[12,200,201,202,205,206,209],{},"L'article de conception présente des expériences préliminaires que les auteurs eux-mêmes demandent de considérer « uniquement comme des points de référence préliminaires » (",[21,203,64],{"href":36,"rel":204,"target":26},[25],"). Chaque essai injectait environ 300 messages de discussion de groupe et de CI GitHub, dont une vingtaine de vraies demandes de travail, et chaque configuration a tourné cinq fois. Le travail de code était remplacé par un outil qui se contentait de consommer du temps : le test mesure donc la coordination, pas la qualité du code, une limite que ",[21,207,44],{"href":42,"rel":208,"target":26},[25]," souligne aussi.",[12,211,212],{},"Dans ce cadre, le Super-agent atteint un temps de réalisation médian (p50) de 12,4 minutes et un p95 de 15,7 minutes. Une session unique avec un outil de sous-agent et un prompt neutre atteint un p95 de 29,5 minutes. L'article l'explique : avec le prompt neutre, le modèle n'a appelé aucun sous-agent sur les cinq essais, si bien que tout le travail est resté dans la session principale, sans cesse interrompue. Avec un prompt incitant à déléguer, il a délégué dans quatre essais sur cinq et a gagné en rapidité, tout en restant globalement plus lent.",[54,214,216],{"id":215},"lire-les-notes-de-sécurité-avant-de-linstaller","Lire les notes de sécurité avant de l'installer",[12,218,14,219,224,225,228],{},[21,220,223],{"href":221,"rel":222,"target":26},"https:\u002F\u002Fgithub.com\u002Fdeepseek-ai\u002Fdsh-superagent-prototype\u002Fblob\u002Fmain\u002Fdocs\u002F006-deployment.md",[25],"guide de déploiement"," est sans détour sur les risques. Les deux presets tournent en ",[159,226,227],{},"danger-full-access"," avec l'utilisateur système de l'opérateur. Un Worker « peut exécuter des programmes par leur chemin et lire le répertoire personnel de l'opérateur », et les wrappers de comptes choisissent les identifiants mais « n'isolent pas le processus ». DeepSeek recommande un compte système dédié, qui ne contient que les identifiants de l'agent.",[12,230,231],{},"Autres prérequis listés dans le même guide :",[88,233,234,237,240,243],{},[91,235,236],{},"un compte utilisateur GitHub distinct pour l'agent, avec un accès en écriture au dépôt (une GitHub App ne peut pas jouer ce rôle, car l'agent lit les notifications du compte) ;",[91,238,239],{},"Node.js 22.19+ ou 24+, pnpm, git 2.32+, Python 3.8+, tmux et GitHub CLI 2.48+ ;",[91,241,242],{},"un dépôt sur github.com, GitHub Enterprise Server n'étant pas pris en charge ;",[91,244,245],{},"un fournisseur de modèle configuré dans DeepSeek Harness ; le bundle utilise le modèle par défaut du profil, sauf indication contraire.",[54,247,249],{"id":248},"pourquoi-cest-intéressant-pour-les-builders","Pourquoi c'est intéressant pour les builders",[12,251,252,253,257],{},"Les agents de code branchés sur GitHub ne sont pas nouveaux : le ",[21,254,256],{"href":42,"rel":255,"target":26},[25],"comparatif de MarkTechPost"," cite l'agent cloud de GitHub Copilot, la GitHub Action de Claude Code et OpenAI Codex. Ce qu'apporte dsh-superagent, c'est une conception publiée et lisible pour faire tourner de nombreuses tâches isolées derrière un point d'entrée conversationnel unique, en traitant l'autorisation et le bruit comme des sujets d'architecture plutôt que comme des réglages de prompt.",[12,259,260],{},"Trois idées restent utiles même sans l'installer :",[262,263,264,270,276],"ol",{},[91,265,266,269],{},[16,267,268],{},"Garder le coordinateur léger."," Un Router doté de peu d'outils et de délais courts reste réactif pendant que les Workers font le gros du travail.",[91,271,272,275],{},[16,273,274],{},"Rendre la délégation structurelle."," La référence testée par DeepSeek montre qu'un modèle peut ne pas déléguer quand le prompt lui laisse le choix. Découper le travail au niveau du système a supprimé cette dépendance dans leur test.",[91,277,278,281,282,285],{},[16,279,280],{},"Limiter l'autorité à chaque tâche."," Chaque Ticket reste limité par son Mandate, et les messages des fils qu'il suit lui parviennent comme contexte mais, selon le ",[21,283,132],{"href":130,"rel":284,"target":26},[25],", n'accordent aucune autorité supplémentaire.",[12,287,288,289,291],{},"Le code est une publication de recherche, en un seul commit, assortie d'un avertissement explicite contre tout usage en production. Pour l'essayer, mieux vaut commencer avec une liste d'autorisation courte, le mode ",[159,290,171],{}," et un compte isolé, comme le conseille la documentation du projet.",{"title":293,"searchDepth":294,"depth":294,"links":295},"",2,[296,297,298,299,300,301],{"id":56,"depth":294,"text":57},{"id":71,"depth":294,"text":72},{"id":123,"depth":294,"text":124},{"id":197,"depth":294,"text":198},{"id":215,"depth":294,"text":216},{"id":248,"depth":294,"text":249},"vibecoding","2026-11-02T20:20:00+01:00","L'équipe Harness de DeepSeek a publié dsh-superagent, un prototype de recherche sous licence MIT qui transforme un dépôt GitHub en agent de code proactif : un Router permanent reçoit les demandes et confie chacune à une session Worker isolée. Fonctionnement, mesures de DeepSeek et points de vigilance avant de l'installer.","md","\u002Fdeepseek-dsh-superagent-open-source-router-worker-coding-agent.jpg",{},true,"\u002Ffr\u002Fnews\u002Fdeepseek-dsh-superagent-agent-de-code-open-source-router-worker",{"title":5,"description":304},"Le prototype dsh-superagent de DeepSeek, sous licence MIT, confie les demandes GitHub à des Workers isolés. Fonctionnement, benchmark et précautions de sécurité.","DeepSeek publie dsh-superagent, un agent de code Router-Worker open source","fr\u002Fnews\u002Fdeepseek-dsh-superagent-agent-de-code-open-source-router-worker",[315,316,317,318,319,320],"deepseek","coding-agents","open-source","github","agent-architecture","dsh-superagent","deepseek-dsh-superagent-open-source-router-worker-coding-agent","ANxUq1eySTRpivRQqDUk7sAgTgJuhrHFTKG2ay5HAMM",[324,331,338],{"title":325,"description":326,"path":327,"date":328,"image":329,"category":330},"ChatGPT et Codex partageront un seul quota d'usage dès décembre : ce qui change pour les développeurs","OpenAI annonce qu'à partir de décembre, les offres Plus, Pro et Business puiseront l'usage de ChatGPT et de Codex dans un même quota mesuré en tokens, hors chat Instant. La taille de ce nouveau quota n'est pas publiée. Ce que l'on sait et comment s'y préparer.","\u002Ffr\u002Fnews\u002Fopenai-chatgpt-codex-quota-partage-decembre-2026","2026-11-02T19:50:00+01:00","\u002Fopenai-chatgpt-codex-shared-usage-allowance-december-2026.jpg","tools",{"title":332,"description":333,"path":334,"date":335,"image":336,"category":337},"S-1 amendé d'Anthropic : 17 milliards de dollars de revenus au T3, et ce que le dossier dit aux développeurs qui construisent sur Claude","Le prospectus d'introduction en bourse mis à jour par Anthropic ajoute les chiffres du troisième trimestre : 17,0 milliards de dollars de revenus, une croissance trimestrielle qui ralentit, un coût du compute remonté à 58 % du chiffre d'affaires et 895,4 milliards d'engagements. Ce que cela change pour les équipes qui dépendent de l'API Claude.","\u002Ffr\u002Fnews\u002Fs1-amende-anthropic-t3-2026-17-milliards-revenus-compute-developpeurs-claude","2026-11-02T18:05:00+01:00","\u002Fanthropic-amended-s1-q3-2026-17-billion-revenue-compute-claude-developers.jpg","ai",{"title":339,"description":340,"path":341,"date":342,"image":343,"category":337},"Hark lève 700 M$ à 6 milliards de valorisation : Brett Adcock veut construire l'interface universelle entre l'humain et l'IA","La startup Hark de Brett Adcock (Figure.AI, Archer Aviation) a levé 700 M$ en Série A à 6 milliards de valorisation le 21 mai 2026. Tour mené par Parkway Venture Capital, avec NVIDIA, AMD Ventures, Salesforce Ventures. Objectif : un appareil hardware natif à l'IA, piloté par speech, vision, mémoire.","\u002Ffr\u002Fnews\u002Fhark-brett-adcock-700-millions-6-milliards-ia-hardware-interface-universelle","2026-05-21T14:00:00","\u002Fimages\u002Fnews\u002Fdefault-og.jpg",{"en":345,"fr":612,"es":7},{"id":346,"title":347,"author":6,"author_avatar":7,"body":348,"category":302,"date":303,"description":603,"extension":305,"image":306,"meta":604,"navigation":308,"path":605,"seo":606,"seoDescription":607,"seoTitle":608,"stem":609,"tags":610,"translationKey":321,"updatedAt":303,"__hash__":611},"news_en\u002Fnews\u002Fdeepseek-dsh-superagent-open-source-router-worker-coding-agent.md","DeepSeek open-sources dsh-superagent, a coding agent that splits routing from execution",{"type":9,"value":349,"toc":595},[350,372,379,383,390,393,397,409,423,431,439,443,451,492,501,505,516,519,523,533,536,550,554,562,565,589],[12,351,352,353,356,357,360,361,33,363,367,368,371],{},"On ",[16,354,355],{},"October 31, 2026",", the DeepSeek Harness team published ",[21,358,27],{"href":23,"rel":359,"target":26},[25]," on GitHub under the MIT license, alongside a design article titled ",[30,362,32],{},[21,364,366],{"href":36,"rel":365,"target":26},[25],"English translation","). By Monday evening the repository had passed 12,000 stars. ",[21,369,44],{"href":42,"rel":370,"target":26},[25]," describes it as a TypeScript bundle, version 0.1.0, released as a single commit.",[12,373,374,375,378],{},"The ",[21,376,51],{"href":23,"rel":377,"target":26},[25]," is explicit about its status: it is a \"research prototype for reference only, not an official DeepSeek product\", much of it was built with coding agents, and it differs from both DeepSeek's internal version and the system described in the article. It is still worth reading for anyone building with coding agents, because it tackles a problem most current tools leave to the user: managing several agent sessions at once.",[54,380,382],{"id":381},"the-problem-sessions-do-not-scale-with-people","The problem: sessions do not scale with people",[12,384,385,386,65],{},"Most coding agents treat one session as the unit of work. In the design article, DeepSeek's team writes that \"one person can usually keep track of only three to five active Sessions at a time\", so the limit on parallel work becomes human attention rather than model capability (",[21,387,389],{"href":36,"rel":388,"target":26},[25],"translation",[12,391,392],{},"Their answer is what they call a Super-agent: one continuous agent that people address from the places they already work, such as GitHub comments or a chat, while the system decides how to split the work behind the scenes. The article lists four properties it aims for: continuity, efficiency, proactivity and persistence.",[54,394,396],{"id":395},"how-the-prototype-works","How the prototype works",[12,398,399,400,403,404,408],{},"dsh-superagent is packaged as a bundle for ",[21,401,80],{"href":78,"rel":402,"target":26},[25]," and watches one GitHub repository, with Feishu as an optional chat source. The ",[21,405,407],{"href":42,"rel":406,"target":26},[25],"architecture has two layers",":",[88,410,411,417],{},[91,412,413,416],{},[16,414,415],{},"A Router"," (the main session) runs continuously, reads incoming events and decides whether they ask for new work. It uses a restricted preset with a smaller tool set.",[91,418,419,422],{},[16,420,421],{},"Workers"," (one per work item, called a Ticket) each get their own agent session and directory, do the actual coding and reply in the original conversation, for example with a pull request.",[12,424,425,426,430],{},"Each Ticket carries an immutable Mandate that records why the work may proceed. Later comments on a pull request reach the right Worker through a Subscription, without anyone having to mention the agent again. According to the ",[21,427,429],{"href":108,"rel":428,"target":26},[25],"execution docs",", foreground shell commands are moved to the background after five seconds, so a Worker running a long build can still receive new messages.",[12,432,433,434,438],{},"A local web panel lists only what needs a human decision. That reflects the rule set out in the ",[21,435,437],{"href":117,"rel":436,"target":26},[25],"philosophy document",": \"Take as little human attention as possible, and give as much value as possible.\" The agent confirms a request once, stays quiet while it works, and writes again only with a result, a question or a failure.",[54,440,442],{"id":441},"proactive-work-is-off-by-default-in-practice","Proactive work is off by default, in practice",[12,444,445,446,450],{},"The agent can take on work nobody explicitly asked for, but only under rules the team writes. The ",[21,447,449],{"href":130,"rel":448,"target":26},[25],"authorization document"," defines three accept modes:",[135,452,453,462],{},[138,454,455],{},[141,456,457,459],{},[144,458,146],{},[144,460,461],{},"Effect",[151,463,464,473,483],{},[141,465,466,470],{},[156,467,468],{},[159,469,161],{},[156,471,472],{},"Nothing inferred becomes a candidate",[141,474,475,480],{},[156,476,477,479],{},[159,478,171],{}," (default)",[156,481,482],{},"Inferred candidates are assessed and positive ones recorded as \"would accept\", but nothing is accepted",[141,484,485,489],{},[156,486,487],{},[159,488,182],{},[156,490,491],{},"A positive assessment accepts the work",[12,493,494,495,497,498,500],{},"Explicit requests from people on the allowed list work in every mode. The docs recommend starting in ",[159,496,171],{}," and switching to ",[159,499,182],{}," only after reviewing what the agent would have done. The same document states that messages, comments and quoted documents \"cannot change\" the permissions enforced by the host, which matters for prompt injection through issue comments.",[54,502,504],{"id":503},"what-deepseek-measured","What DeepSeek measured",[12,506,507,508,511,512,515],{},"The design article includes preliminary experiments that the authors themselves say \"should be treated only as preliminary reference points\" (",[21,509,389],{"href":36,"rel":510,"target":26},[25],"). Each run fed about 300 group-chat and GitHub CI messages, about 20 of them real work requests, and each setup ran five times. Coding work was replaced by a tool that only consumed time, so the test measures coordination, not code quality, a limit ",[21,513,44],{"href":42,"rel":514,"target":26},[25]," also points out.",[12,517,518],{},"In that setting, the Super-agent reached a p50 task completion time of 12.4 minutes and a p95 of 15.7 minutes. A single session with a subagent tool and a neutral prompt reached a p95 of 29.5 minutes. The article explains why: under the neutral prompt, the model did not call a subagent in any of the five runs, so all work stayed in the main session and kept getting interrupted. With a delegation-first prompt, it delegated in four of five runs and got faster, but remained slower overall.",[54,520,522],{"id":521},"read-the-security-notes-before-you-run-it","Read the security notes before you run it",[12,524,374,525,529,530,532],{},[21,526,528],{"href":221,"rel":527,"target":26},[25],"deployment guide"," is direct about the risks. Both presets run with ",[159,531,227],{}," as the operator's OS user. A Worker \"can execute programs by path and read the operator's home directory\", and the account wrappers choose credentials but \"do not isolate the process\". DeepSeek recommends a dedicated OS account that holds only the agent's credentials.",[12,534,535],{},"Other requirements from the same guide:",[88,537,538,541,544,547],{},[91,539,540],{},"a separate GitHub user account for the agent with write access to the repository (a GitHub App cannot be the agent, because it reads the account's notifications);",[91,542,543],{},"Node.js 22.19+ or 24+, pnpm, git 2.32+, Python 3.8+, tmux and GitHub CLI 2.48+;",[91,545,546],{},"a repository on github.com, since GitHub Enterprise Server is not supported;",[91,548,549],{},"a model provider configured for DeepSeek Harness; the bundle uses the profile's default model unless told otherwise.",[54,551,553],{"id":552},"why-it-matters-for-builders","Why it matters for builders",[12,555,556,557,561],{},"Coding agents on GitHub are not new: ",[21,558,560],{"href":42,"rel":559,"target":26},[25],"MarkTechPost's comparison"," lists GitHub Copilot's cloud agent, the Claude Code GitHub Action and OpenAI Codex. What dsh-superagent adds is a published, readable design for running many isolated work items behind a single conversational entry point, with authorization and noise control treated as first-class concerns rather than prompt tweaks.",[12,563,564],{},"Three ideas carry over even if you never install it:",[262,566,567,573,579],{},[91,568,569,572],{},[16,570,571],{},"Keep the coordinator light."," A router with a small tool set and short timeouts stays responsive while workers do the heavy lifting.",[91,574,575,578],{},[16,576,577],{},"Make delegation structural."," DeepSeek's own baseline shows a model may not delegate when the prompt leaves it the choice. Splitting work at the system level removed that dependency in their test.",[91,580,581,584,585,588],{},[16,582,583],{},"Scope authority per task."," Each Ticket stays limited by its Mandate, and messages on threads it follows reach it as context but, per the ",[21,586,449],{"href":130,"rel":587,"target":26},[25],", grant no new authority.",[12,590,591,592,594],{},"The code is a research release with one commit and an explicit warning against treating it as production software. Anyone trying it should start with a short allowed list, ",[159,593,171],{}," mode and an isolated account, as the project's own documentation advises.",{"title":293,"searchDepth":294,"depth":294,"links":596},[597,598,599,600,601,602],{"id":381,"depth":294,"text":382},{"id":395,"depth":294,"text":396},{"id":441,"depth":294,"text":442},{"id":503,"depth":294,"text":504},{"id":521,"depth":294,"text":522},{"id":552,"depth":294,"text":553},"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",{"title":347,"description":603},"DeepSeek's MIT-licensed dsh-superagent prototype routes GitHub requests to isolated Worker sessions. How it works, the benchmark and the security caveats.","DeepSeek open-sources dsh-superagent, a Router-Worker coding agent","news\u002Fdeepseek-dsh-superagent-open-source-router-worker-coding-agent",[315,316,317,318,319,320],"YNP-SBEnjdwaLGYiyrRCSwr_F1G_7kewpWqRF7ZjWOw",{"id":4,"title":5,"author":6,"author_avatar":7,"body":613,"category":302,"date":303,"description":304,"extension":305,"image":306,"meta":794,"navigation":308,"path":309,"seo":795,"seoDescription":311,"seoTitle":312,"stem":313,"tags":796,"translationKey":321,"updatedAt":303,"__hash__":322},{"type":9,"value":614,"toc":786},[615,630,635,637,642,644,646,654,664,669,674,676,681,717,723,725,733,735,737,744,746,756,758,763,765,782],[12,616,14,617,19,619,28,622,33,624,39,627,45],{},[16,618,18],{},[21,620,27],{"href":23,"rel":621,"target":26},[25],[30,623,32],{},[21,625,38],{"href":36,"rel":626,"target":26},[25],[21,628,44],{"href":42,"rel":629,"target":26},[25],[12,631,14,632,52],{},[21,633,51],{"href":23,"rel":634,"target":26},[25],[54,636,57],{"id":56},[12,638,60,639,65],{},[21,640,64],{"href":36,"rel":641,"target":26},[25],[12,643,68],{},[54,645,72],{"id":71},[12,647,75,648,81,651,86],{},[21,649,80],{"href":78,"rel":650,"target":26},[25],[21,652,85],{"href":42,"rel":653,"target":26},[25],[88,655,656,660],{},[91,657,658,96],{},[16,659,95],{},[91,661,662,102],{},[16,663,101],{},[12,665,105,666,111],{},[21,667,110],{"href":108,"rel":668,"target":26},[25],[12,670,114,671,120],{},[21,672,119],{"href":117,"rel":673,"target":26},[25],[54,675,124],{"id":123},[12,677,127,678,133],{},[21,679,132],{"href":130,"rel":680,"target":26},[25],[135,682,683,691],{},[138,684,685],{},[141,686,687,689],{},[144,688,146],{},[144,690,149],{},[151,692,693,701,709],{},[141,694,695,699],{},[156,696,697],{},[159,698,161],{},[156,700,164],{},[141,702,703,707],{},[156,704,705,172],{},[159,706,171],{},[156,708,175],{},[141,710,711,715],{},[156,712,713],{},[159,714,182],{},[156,716,185],{},[12,718,188,719,191,721,194],{},[159,720,171],{},[159,722,182],{},[54,724,198],{"id":197},[12,726,201,727,205,730,209],{},[21,728,64],{"href":36,"rel":729,"target":26},[25],[21,731,44],{"href":42,"rel":732,"target":26},[25],[12,734,212],{},[54,736,216],{"id":215},[12,738,14,739,224,742,228],{},[21,740,223],{"href":221,"rel":741,"target":26},[25],[159,743,227],{},[12,745,231],{},[88,747,748,750,752,754],{},[91,749,236],{},[91,751,239],{},[91,753,242],{},[91,755,245],{},[54,757,249],{"id":248},[12,759,252,760,257],{},[21,761,256],{"href":42,"rel":762,"target":26},[25],[12,764,260],{},[262,766,767,771,775],{},[91,768,769,269],{},[16,770,268],{},[91,772,773,275],{},[16,774,274],{},[91,776,777,281,779,285],{},[16,778,280],{},[21,780,132],{"href":130,"rel":781,"target":26},[25],[12,783,288,784,291],{},[159,785,171],{},{"title":293,"searchDepth":294,"depth":294,"links":787},[788,789,790,791,792,793],{"id":56,"depth":294,"text":57},{"id":71,"depth":294,"text":72},{"id":123,"depth":294,"text":124},{"id":197,"depth":294,"text":198},{"id":215,"depth":294,"text":216},{"id":248,"depth":294,"text":249},{},{"title":5,"description":304},[315,316,317,318,319,320]]