iFANN
    Rechercher sur iFANN...
    Connexion
    Accueil
    Actualités
    Vidéos
    Photos
    GIF
    Explorer
    Sondages
    Récompenses
    iFAMOUS
    Wiki
    Animé
    Salons
    Notifications
    Messages
    Favoris
    Profil
    WikiRécompensesiFAMOUSClassementsSecteursRécompenses créateursRécompenses utilisateursConditionsConfidentialitéRègles de la communautéRetrait / DMCAAideDéveloppeurs

    © 2026 iFANN

    Accueil
    Rechercher
    Messages
    Alertes
    Profil
    Photo
    Vojtech
    Vojtech@vojtech3mo
    🏢Microsoft🏛️MIT - Massachusetts Institute of Technology📱Kimi K3
    Loop vs graph agents explained

    @vojtechYour agent keeps forgetting everything between runs, and that's the real bottleneck. A loop-based setup gives the agent a goal and lets it figure things out, which works for fresh problems but falls apart when you repeat the same task. A graph instead structures the work as connected facts the agent moves through, so knowledge builds up instead of resetting to zero. It's more effort upfront, but you never start from nothing again. The loop forgets, the graph accumulates. That's the actual unlock: organize what you know as relationships rather than loose files, and a smaller model moving through a clean graph beats a bigger one searching through raw text. Microsoft, Stanford, and MIT all confirmed this independently. Kimi K3 fits the role: a million tokens of context can handle an entire reasoning path at once, at a third of the cost of frontier models. When a loop stops teaching you anything new, freeze it into a graph. So which one is your stack actually running?

    Voir la publication d'origine

    Loop vs graph agents explained

    Photo de @vojtech· Aug 16, 2026· Microsoft

    À propos de cette photo

    This is a diagram illustrating a process. The diagram is divided into three sections: "Extract - what gets stored?", "Retrieve - how do you find things?", and "Loop - how does it compound?". A final section titled "Putting it all together" shows a workflow from "Question" to "New facts". The style is clean and informative, like a technical explanation. The text overlays are the titles and descriptions within the diagram boxes.

    Voir toutes les photos de MicrosoftLire le wiki de Microsoft

    ?

    Aucun commentaire pour le moment. Soyez le premier !

    Plus de photos de Microsoft

    Voir toutes les photos de Microsoft
    Satya Nadella speaking at stage event2Satya Nadella speaking at stage eventUS flight suspensions and 2026 Friday the 13th2US flight suspensions and 2026 Friday the 13thMicrosoft internet issue clarificationMicrosoft internet issue clarificationJapan stock market fax machines2Japan stock market fax machinesMicrosoft Outlook Teams outage2Microsoft Outlook Teams outageMicrosoft outage delays US jail releasesMicrosoft outage delays US jail releasesMicrosoft outage continuesMicrosoft outage continuesMicrosoft Outlook outageMicrosoft Outlook outageOutlook service statusOutlook service statusMicrosoft patent pays users for idle GPU AI training2Microsoft patent pays users for idle GPU AI trainingMicrosoft market cap surpasses Nvidia2Microsoft market cap surpasses NvidiaUS stock market gainUS stock market gainSatya Nadella on AI security2Satya Nadella on AI securityMicrosoft Azure 48% revenue jumpMicrosoft Azure 48% revenue jumpCall of Duty: Modern Warfare 4 Xbox reviewsCall of Duty: Modern Warfare 4 Xbox reviewsMicrosoft AI portfolio ranked by valuationMicrosoft AI portfolio ranked by valuationMicrosoft revenue vs headcount AIMicrosoft revenue vs headcount AIMicrosoft 27% workforce cutMicrosoft 27% workforce cut
    Photo
    Vojtech
    Vojtech@vojtech3mo
    🏢Microsoft🏛️MIT - Massachusetts Institute of Technology📱Kimi K3
    Loop vs graph agents explained

    @vojtechYour agent keeps forgetting everything between runs, and that's the real bottleneck. A loop-based setup gives the agent a goal and lets it figure things out, which works for fresh problems but falls apart when you repeat the same task. A graph instead structures the work as connected facts the agent moves through, so knowledge builds up instead of resetting to zero. It's more effort upfront, but you never start from nothing again. The loop forgets, the graph accumulates. That's the actual unlock: organize what you know as relationships rather than loose files, and a smaller model moving through a clean graph beats a bigger one searching through raw text. Microsoft, Stanford, and MIT all confirmed this independently. Kimi K3 fits the role: a million tokens of context can handle an entire reasoning path at once, at a third of the cost of frontier models. When a loop stops teaching you anything new, freeze it into a graph. So which one is your stack actually running?

    Voir la publication d'origine

    Loop vs graph agents explained

    Photo de @vojtech· Aug 16, 2026· Microsoft

    À propos de cette photo

    This is a diagram illustrating a process. The diagram is divided into three sections: "Extract - what gets stored?", "Retrieve - how do you find things?", and "Loop - how does it compound?". A final section titled "Putting it all together" shows a workflow from "Question" to "New facts". The style is clean and informative, like a technical explanation. The text overlays are the titles and descriptions within the diagram boxes.

    Voir toutes les photos de MicrosoftLire le wiki de Microsoft

    ?

    Aucun commentaire pour le moment. Soyez le premier !

    Plus de photos de Microsoft

    Voir toutes les photos de Microsoft
    Satya Nadella speaking at stage event2Satya Nadella speaking at stage eventUS flight suspensions and 2026 Friday the 13th2US flight suspensions and 2026 Friday the 13thMicrosoft internet issue clarificationMicrosoft internet issue clarificationJapan stock market fax machines2Japan stock market fax machinesMicrosoft Outlook Teams outage2Microsoft Outlook Teams outageMicrosoft outage delays US jail releasesMicrosoft outage delays US jail releasesMicrosoft outage continuesMicrosoft outage continuesMicrosoft Outlook outageMicrosoft Outlook outageOutlook service statusOutlook service statusMicrosoft patent pays users for idle GPU AI training2Microsoft patent pays users for idle GPU AI trainingMicrosoft market cap surpasses Nvidia2Microsoft market cap surpasses NvidiaUS stock market gainUS stock market gainSatya Nadella on AI security2Satya Nadella on AI securityMicrosoft Azure 48% revenue jumpMicrosoft Azure 48% revenue jumpCall of Duty: Modern Warfare 4 Xbox reviewsCall of Duty: Modern Warfare 4 Xbox reviewsMicrosoft AI portfolio ranked by valuationMicrosoft AI portfolio ranked by valuationMicrosoft revenue vs headcount AIMicrosoft revenue vs headcount AIMicrosoft 27% workforce cutMicrosoft 27% workforce cut