{"article_id":"2736","schema_version":2,"source_content_hash":"effa3ec2226ff75e1a8bea8d5190c514393e51c8a812bbec3ba2fabb16babea0","computed_at":"2026-07-12T12:30:33Z","payload":{"topic":{"subtopic":"Архитектура и ограничения агентов","canonical":"Разница между LLM и AI-агентами","confidence":0.95,"related_topics":["архитектура агентов","безопасность AI-агентов","OpenClaw","ограничения контекста"]},"entities":[{"role":"primary","type":"technology","canonical_name":"LLM"},{"role":"primary","type":"technology","canonical_name":"AI-агент"},{"role":"primary","type":"technology","canonical_name":"OpenClaw"},{"role":"secondary","type":"technology","canonical_name":"Claude Code"},{"role":"secondary","type":"technology","canonical_name":"Cursor"},{"role":"secondary","type":"technology","canonical_name":"Devin"},{"role":"secondary","type":"company","canonical_name":"Anthropic"},{"role":"secondary","type":"organization","canonical_name":"Censys"},{"role":"secondary","type":"organization","canonical_name":"Tenzai"},{"role":"secondary","type":"person","canonical_name":"Саймон Уиллисон"}],"passages":[{"kind":"definition","text":"Обычная LLM — это просто генератор текста. Вы даёте prompt, она даёт ответ. Весь её мир — это окно контекста в 128–200 тысяч токенов. Она не помнит, что было вчера. Она не может сама решить, что ей нужно сделать. Она реактивна — ждёт, пока её дёрнут. Агент — это система, у которой есть память, планировщик, инструменты и цикл обратной связи."},{"kind":"comparison","text":"Если обычная LLM скажет вам: «Чтобы забронировать столик, откройте сайт ресторана», то агент OpenClaw сам откроет браузер, зайдёт на нужный сайт, заполнит форму, подтвердит бронь и вернёт вам скриншот с подтверждением."},{"kind":"answer","text":"Главное заблуждение: «Агент — это LLM, которая умеет вызывать функции». На деле архитектура агента — это многослойный пирог: ядро (LLM), планировщик, инструменты (клешни), память и шлюз."},{"kind":"definition","text":"Контекстная деградация — это бич всех агентов. LLM не умеет держать в голове 200 тысяч токенов так же хорошо, как первые 10 тысяч. Чем длиннее сессия, тем ниже качество решений. Агент может «забыть» первые шаги плана, начать ходить по кругу или потерять нить задачи."},{"kind":"troubleshooting","text":"Если вы используете OpenClaw — убедитесь, что шлюз не торчит наружу, настройте минимальные права для инструментов и следите за расходом токенов."},{"kind":"how_to","text":"Если вы выбираете между Cursor и Claude Code — поймите, что первый лучше для интерактивной работы (автодополнение, быстрые правки в IDE), второй — для автономных многошаговых задач (запустить тесты, пофиксить баги, закоммитить)."}],"faq_items":[{"answer":"Обычная LLM — это генератор текста, реагирующий на запросы. Агент имеет память, планировщик, инструменты и цикл обратной связи, позволяющий самостоятельно выполнять многошаговые задачи.","question":"Чем AI-агент отличается от обычной LLM?","confidence":0.95},{"answer":"С ростом длины сессии LLM хуже удерживает информацию в контексте. После 30-40 шагов агент может начать игнорировать промежуточные результаты, ходить по кругу или терять нить задачи.","question":"Что такое контекстная деградация у агентов?","confidence":0.95},{"answer":"Один запрос может потребовать 5-10 обращений к LLM: разбор намерения, планирование, выполнение инструментов, обработка ошибок и подтверждение. OpenClaw тратит в 3-5 раз больше токенов, чем ручной режим.","question":"Почему AI-агенты потребляют много токенов?","confidence":0.95},{"answer":"Агенты с доступом к файлам, коду и интернету образуют «смертельную тройку». Более 21 000 экземпляров OpenClaw были публично доступны без пароля. Также встречаются вредоносные расширения, ворующие криптокошельки и API-ключи.","question":"Какие основные проблемы безопасности у агентов?","confidence":0.95}],"seo_title":"Разница между LLM и AI-агентом: архитектура, ограничения, безопасность","ai_summary":"Статья объясняет, что AI-агент — это не просто LLM с функциями, а сложная архитектура с памятью, планировщиком, инструментами и шлюзом. На примере OpenClaw разбираются компоненты агента и его ограничения: контекстная деградация, высокий расход токенов, проблемы безопасности и отсутствие бизнес-логики. Даются рекомендации по выбору агента и настройке безопасности.","meaning_text":"Центральная мысль: AI-агент — это не просто расширенная LLM, а принципиально иная архитектура с собственными компромиссами; успешное использование агентов требует понимания их внутреннего устройства и ограничений.","subject_text":"Статья посвящена сравнению LLM и AI-агентов, подробно разбирает архитектуру агентов на примере OpenClaw и описывает их основные ограничения и практические рекомендации.","key_takeaways":["AI-агент отличается от LLM наличием памяти, планировщика, инструментов и цикла обратной связи.","Архитектура агента включает ядро (LLM), планировщик, инструменты (клешни), память и шлюз.","Главные ограничения агентов: контекстная деградация, высокий расход токенов, уязвимости безопасности и отсутствие здравого смысла в бизнес-логике.","При выборе агента учитывайте задачу: Cursor для интерактивной работы, Claude Code для автономных многошаговых задач.","Для безопасного использования настройте минимальные права, не оставляйте шлюз открытым и следите за расходом токенов."],"author_profile":{"bio":"Эксперт в области автомобилестроения и ИИ","content_centroid":[-0.028203053,-0.015130794,-0.058882188,0.041680384,-0.01176186,0.021584362,-0.030592432,-0.009055316,0.009921041,0.019872187,0.0009997953,0.0048083044,-0.019446475,-0.020689417,-0.018708868,-0.053409386,0.000042274543,-0.0031215586,-0.0046868594,-0.023739496,0.02273863,0.01879179,-0.025759421,-0.011750937,0.017624168,-0.031006735,-0.011699744,-0.015503349,0.03393645,0.01882408,0.014204289,0.007956676,0.013522016,-0.037204247,-0.024902422,0.022222042,-0.033216134,-0.04130779,0.0005401041,0.00792252,-0.0024454861,0.030241331,0.0015145579,-0.063938454,-0.04453313,-0.052848626,0.0039471528,-0.023901133,-0.032367535,-0.0008061604,0.0019428203,0.048155334,-0.020967392,-0.019396309,-0.0010717642,0.0035191022,-0.017238583,-0.01011866,-0.028649403,0.015251758,-0.041396424,0.0055706995,-0.05313706,0.049938157,0.037473783,0.065968,-0.032412533,-0.02286164,0.0013123002,-0.032120712,-0.0047379206,-0.066396005,-0.01913617,-0.0091094375,0.029203573,0.038019925,0.016989868,0.008034153,0.036410525,-0.033346172,0.009062272,0.02070364,-0.039707676,-0.059909347,-0.04981023,0.027658015,-0.040174346,-0.016767839,-0.026130857,-0.015781721,-0.009783408,0.022602176,0.01333951,-0.0033854044,0.0026495273,0.02093928,0.0050796084,0.012789754,-0.03244635,0.012603685,0.0106931105,0.030099455,-0.029160133,0.038254432,0.016591962,-0.02007593,-0.009496299,0.0046238075,-0.020895002,0.004632505,-0.0042762333,-0.065706454,-0.035453312,-0.009685165,0.060237702,-0.019914247,0.011565511,-0.007212054,0.0065692933,-0.0069039282,-0.0074454546,-0.02374742,0.023201684,0.016379,-0.05434705,-0.04918702,-0.020569023,0.03597595,-0.045880444,0.009396827,0.004392934,0.027522769,-0.020971615,-0.02906288,-0.047661856,0.01377366,0.05198653,0.026267963,-0.006429632,0.08952453,0.009610935,-0.034294087,-0.03418558,0.07596036,0.023842173,0.029545635,0.019015402,0.0043302663,-0.04889829,-0.013211136,0.00091930816,0.0020758484,-0.055420842,-0.014835247,0.03688502,-0.011196661,0.00071038556,-0.009693848,0.0032238257,0.029214056,-0.007335248,0.008577038,-0.045413446,-0.036989845,-0.00055767066,-0.0064546457,0.06299787,0.027264556,-0.07028337,-0.0056900126,-0.021848343,-0.017370882,0.0011325104,0.015865378,-0.014050192,-0.02207323,-0.0102720335,0.044070575,0.07473983,-0.020156024,-0.034097694,0.031931285,-0.026166156,-0.043283433,0.009923731,0.07417349,0.04214236,-0.022201668,-0.016681913,0.035287887,0.01319306,0.022225324,-0.073601104,-0.011744114,-0.009117946,0.00806104,-0.021463413,-0.052991766,0.0071920254,-0.010754035,-0.049194537,0.011599259,0.01226954,0.0055621015,-0.0012320854,-0.010264652,0.05395934,0.013363373,0.0082782,-0.009611414,-0.0038894746,-0.012076103,-0.033515688,-0.027260387,0.028614007,-0.049935166,-0.020566307,-0.070166364,0.035104614,-0.016835388,-0.03759095,0.015303938,-0.03761089,-0.0044286507,-0.03561371,-0.022780074,-0.049140204,-0.03768634,0.013876397,-0.015593589,-0.010402648,-0.019699786,0.005044316,-0.0059232004,-0.028809652,0.022244995,-0.008908373,0.0101888385,-0.026303038,-0.0026291343,0.070259295,-0.024016168,-0.028709143,0.02977935,0.054592162,-0.02168626,0.016781004,-0.0017947197,-0.018232346,-0.05926832,0.008344536,-0.0092752455,0.011564059,0.015660783,-0.006416225,0.013002055,0.0024838357,0.002297132,0.00042776813,-0.010163632,-0.04499261,-0.03957232,0.028567037,0.041514117,0.017416447,0.035434216,-0.018314952,-0.028365863,0.0021739416,0.031039653,0.018397685,-0.009582387,-0.043412093,0.008858614,0.043495823,0.031889267,-0.04728061,-0.007955737,0.008170725,0.018047268,0.024433443,0.03538172,0.045687027,0.02603672,0.005971804,0.031697646,-0.011360923,-0.030582473,0.031470604,-0.0002409612,-0.027164178,-0.01913336,-0.0014049327,-0.0084184185,0.0066477577,-0.032952424,-0.055699255,0.039259292,0.031234816,0.0485146,-0.0126710385,-0.040913742,0.021055361,-0.02989798,0.023129301,0.025956003,-0.03246276,0.032642446,-0.0025670864,0.033624277,0.04535854,0.02645489,0.033989258,-0.01976278,0.029204443,-0.035842717,0.007756879,0.006295478,0.033584315,-0.058561407,0.011787458,-0.04181847,0.014777187,0.023498852,-0.03995196,0.037305165,-0.049569692,-0.05215271,0.020974599,0.014693568,-0.037340123,0.0065733436,0.0076903077,-0.01767334,0.008025055,-0.030496756,0.017451415,0.015537617,-0.018193537,0.000675295,-0.0059167203,0.080781765,0.016888665,-0.025682645,0.05372934,0.0047992826,-0.014326254,0.02921459,-0.017536728,-0.05716984,0.0010303314,-0.006399967,-0.06308891,-0.063988194,0.03356534,-0.043945216,0.0009681107,-0.0210154,-0.01801407,0.015227654,-0.033725876,0.03915768,-0.014923438,-0.085075885,-0.030758265,0.019651882,-0.034821622,0.0087476475,0.063399866,-0.008777042,0.002707127,0.04289327,-0.04068716,0.015532874,-0.023333011,-0.028876694,-0.0033947378,-0.008540713,0.0298552,0.03180912,-0.019464102,0.001998344,-0.0073989728,-0.050361194,-0.0548011,-0.019377986,-0.037075248,-0.035065502,-0.015797915,-0.0037616333,-0.019355534,-0.0665567,0.023027182,0.031709787,0.007899247,0.0029146685,0.01067726,0.030208066,0.0035559179,-0.0074167443,0.02324053,-0.030702073,-0.03567769,0.020876966,0.046808142,0.0072451667,0.028526833,-0.042441532,0.015832312,-0.023777029,-0.002657879,-0.025654718,-0.033386305,0.005079709,0.014061338,0.021069333,-0.0154128345,-0.0012872746,-0.00042906532,-0.009205292,-0.082729645,-0.011536452,0.030964978,0.034452952,-0.010375142,-0.056551505,-0.0648419,-0.0034992658,0.04909009,-0.0025348046,-0.012229681,-0.001752134,0.01044567,0.0126998015,-0.007044079,-0.01358285,0.02351054,-0.018036867,0.012336862,0.01436503,0.07470687,0.011927598,-0.014355121,-0.038742155,-0.038099095,0.011139721,-0.033158083,-0.04367094,0.058785036,0.07017903,-0.03048778,0.06514195,0.022874888,0.06153628,0.03579377,-0.009709417,0.014302541,-0.042754516,-0.03723667,-0.039996874,-0.054823518,0.015487272,0.00010255294,0.04255538,0.07664503,-0.0090555865,-0.016168686,-0.013329658,0.046036776,-0.027977591,0.01763102,-0.015140638,0.026258485,0.020994877,-0.012498357,0.007265319,0.037015166,0.022106785,-0.022236524,-0.012943184,0.0026172304,-0.016053868,-0.038790885,-0.07620452,-0.0241411,0.041250892,0.00013373508,0.05749078,-0.024111874,-0.011180726,-0.034672987,0.019773657,0.018286625,-0.03542592,0.048474,0.001474125,-0.0016895164,-0.016536068,-0.0062725125,0.028172174,0.028368441,-0.033935636,0.000051370516,0.007729535,-0.010740853,-0.04273,0.013378525,0.0028246746,0.016371867,0.020494843,-0.026552532,-0.009042672,0.0065705557,-0.022436017,-0.014041901,-0.027120508,0.021591337,-0.027322087,0.023907892,-0.016237982,-0.02652487,-0.0015660488,-0.062337562,-0.0791705,0.03870597,-0.00021006774,0.038532548,0.043889977,-0.027422603,-0.053393833,-0.010079525,0.0000010265711,-0.047722112,0.0010449458,-0.017061623,0.046905145,-0.024478732,0.010231658,0.045773104,0.041355643,0.0018666337,-0.0058746566,0.039236486,0.0142085515,0.013326169,-0.028546933,0.0010152846,0.05805411,0.023740502,-0.029167611,-0.017751755,-0.024504613,-0.014848064,-0.020847147,0.024744319,0.0038061782,0.034587864,0.029565686,-0.0036201598,0.004062837,0.0064057736,-0.018590333,0.004170799,-0.027684495,0.03633055,-0.045978196,0.044360384,0.029037159,0.00981518,-0.012272423,-0.026093436,-0.046808824,-0.027139127,0.028067676,-0.011086896,-0.056147493,0.018972047,0.037983306,-0.025815047,-0.036838308,0.005113991,-0.013361099,0.0044919113,-0.0022091186,-0.01885388,-0.018930184,-0.01002582,-0.06064889,0.018899538,-0.020620355,-0.028375115,0.019059546,-0.00077989296,0.024672542,-0.0004735242,0.054991193,-0.043742042,0.010299481,0.011730002,-0.009968045,-0.029024774,0.010731484,-0.024563866,0.025058713,0.025938991,0.0061272224,0.05662435,-0.022016183,-0.058004662,0.00708347,-0.0077716303,0.04396165,0.012920498,0.01464875,0.03969599,0.010313144,-0.05258085,0.005004895,-0.03435485,-0.024037614,-0.026607228,-0.043365743,-0.0054760426,-0.02353659,-0.0880575,-0.057041626,-0.0000109189195,0.0125801535,0.010214366,0.047456563,0.003857098,-0.0094304,-0.03178968,-0.0126499785,-0.024933329,-0.06964058,-0.027609715,0.035082884,0.03731407,0.034594517,0.002837247,0.03887509,-0.02809319,-0.0035648495,0.032536764,-0.031328376,-0.08189824,-0.03774191,-0.010229299,-0.008805434,0.026864987,-0.024057437,-0.028368598,-0.025856964,0.041729543,0.015504856,0.023462594,-0.02454118,-0.054792956,0.019470658,0.04380285,0.009272702,0.031227874,-0.07719412,-0.031892803,-0.06138462,0.031974815,-0.049787946,0.026321212,0.015614345,0.009001464,-0.011452172,-0.028121838,0.0017413709,-0.06759678,-0.039439056,0.026175834,0.047870465,-0.03656063,-0.014986302,-0.041066572,0.045879077,-0.022147251,-0.02330222,0.03416336,0.00014867449,0.01531133,0.060822517,0.027494414,0.022935707,-0.0069223344,0.019380433,-0.05501665,-0.03952801,-0.047035806,-0.04861282,0.046772283,-0.049686655,-0.03367926,0.03696128,-0.020224648,-0.0045195497,0.004319197,0.019764356,0.009229355,-0.023930984,-0.0023972467,-0.002198025,-0.0059138397,0.00672413,-0.046599194,-0.037699543,0.027264027,0.007336603,-0.015912326,0.03206385,-0.0056790905,-0.014211169,-0.03516951,0.041181933,-0.03715477,0.03185292,-0.010213891,0.0017058435,0.015560411,0.021270461,-0.004718678,-0.019854488,-0.0026164602,-0.035940133,-0.0007028167,0.032428175,0.031326942,0.016947418,0.0093562985,0.03438153,-0.03742286,-0.044334743,0.043230396,0.03812357,-0.016946584,-0.01059209,0.03560896,0.027756432,-0.030695118,-0.02039178,-0.018178463,0.026684964,0.0125823235,0.051591948,0.041577544,-0.009708173,0.016545016,-0.0006224146,-0.037408423,0.05962241,-0.029730048,0.0028849454,-0.06237728,0.025816534,0.020899015,-0.007993392,-0.008363229,0.06835679,-0.015863335,-0.031484254,0.029556248,-0.046708774,-0.012310485,-0.037937496,-0.012685731,0.042244956,0.03169073,0.009014587,-0.010284901,0.0273301,-0.026811965,0.03277251,0.041534085,0.025382366,0.009127135,0.030360684,0.07043198,-0.022782315,0.020335227,-0.0035396637,-0.053410236,0.08359497,0.023419704,0.00097165565,0.044361867,-0.06885962,-0.02053427,-0.02784072,0.030845845,-0.07680844,-0.0071619954,0.03910568,0.007536443,0.033283852,-0.0016232358,-0.018878115,-0.019748505,0.02532278,-0.036291275,0.026416086,-0.024586005,-0.029842243,-0.015004646,0.05010325,-0.0014682001,-0.0391612,-0.03340583,-0.012094342,-0.057963774,0.04935264,-0.050011456,0.021682454,-0.0125371935,-0.015083568,-0.03105225,-0.03369556,-0.01556073,0.050429616,-0.004986362,-0.0046492885,0.054477684,-0.053908963,0.00830141,0.022800235,0.011591314,0.005084632,0.0133301485,0.041691154,0.027099011,0.06263965,0.07205204,0.007331814,-0.023463754,-0.010329976,-0.043690514,-0.039238673,0.05948979,-0.0053297468,-0.004447717,0.02283998,-0.034612637,0.005255315,0.03379213,-0.042071234,-0.026522504,-0.019058345,-0.03198754,0.014036193,-0.013226083,0.0015095889,-0.03326929,-0.062002342,0.044750977,-0.0029533748,0.053940933,-0.009265208,-0.023079015,0.01849856,-0.071004204,0.019387363,0.041019525,-0.03689383,0.035353065,0.0021138024,0.05329965,0.020099822,0.011946386,0.010036341,0.0121527435,0.0506776,0.0415344,0.030484997,0.02693614,-0.08147922,0.009771823,0.06490582,0.05594499,0.008201778,-0.049537394,0.04077235,0.023553716,-0.0407381,-0.04686094,-0.019229777,-0.04071516,-0.01416552,0.02558723,-0.06378933,-0.02573474,0.016069716,-0.021100601,-0.002290412,0.00067508017,-0.020429116,0.021707576,0.017441453,-0.038237754,-0.0465527,0.024415,-0.008047763,0.0095013,-0.04073009,-0.042938534,0.02492264,0.022608135,-0.018378569,-0.0037634452,0.012603653,0.022320503,-0.021851253,0.015764525,0.038591262,0.014167026,-0.021307321,0.007829513,0.034894533,0.010658232,0.041906267,0.01606578,-0.021822782,0.019987695,0.011102709,0.03998107,-0.02595219,-0.011128655,0.009947896,0.045113426,-0.03479159,-0.012923687,0.01865947,0.009735311,0.03028518,0.0017793556,-0.010112785,0.0445218,0.016328583,-0.023788266,0.035484944,-0.040124886,0.0269189,0.013497449,-0.040966213,0.05384345,-0.0151887415,-0.01810516,0.029302806,0.018818026,-0.035103995,0.01575693,-0.01161545,-0.0018283685,-0.010911692,-0.011754214,-0.0019459861,0.012768472,0.019525914,-0.01642452,0.023799066,0.0047203535,-0.024860073,0.022231786,-0.023672525,0.022780191,0.04127709,-0.03292467,0.023977261,0.07789369,-0.003595899,-0.033932216,-0.015916971,-0.006297606,0.022660116,-0.03631265,-0.026295025,0.01231522,0.009609874,-0.0055371197,0.017973874,0.027636504,-0.031458367,-0.019401828,-0.024980022,0.0346767,-0.011441917,-0.0138230845,0.039831005,0.015776614,0.027867371,-0.008731221,0.0024753718,-0.013531108,-0.004906471,0.031445548,-0.029891096,-0.08632496,0.008080459,0.06018181,0.0010264637,0.0072507206,0.015768446,-0.01978106,0.031339694,0.06029089,0.019994762,0.010099994,0.030118432,-0.008666726,0.031238284,0.0014034646,-0.02081493,-0.042919908,0.046245232,0.02779768,0.016119301,0.03851391,-0.044385016,-0.025656285,0.009054471],"sampled_articles":26},"link_candidates":[{"phrase":"OpenClaw","relation":"инструмент"},{"phrase":"Claude Code","relation":"инструмент"},{"phrase":"Cursor","relation":"инструмент"},{"phrase":"Devin","relation":"инструмент"},{"phrase":"контекстная деградация","relation":"термин"},{"phrase":"безопасность агентов","relation":"тема"}],"seo_description":"Статья объясняет принципиальную разницу между обычной LLM и AI-агентом, разбирает архитектуру агентов на примере OpenClaw, описывает ограничения (контекстная деградация, расход токенов, безопасность) и даёт рекомендации по выбору и использованию. Узнайте, чем агент отличается от чата, какие подводные камни ждут и как…","experience_labels":["explanation","new_perspective","practical_advice","investigation"],"overall_confidence":0.95}}