Umhlahlandlela wonjiniyela wezinqubo zokwenziwa kwe-AI kanye nezakhiwo

Ikhaya » Python » Umhlahlandlela wonjiniyela wezinqubo zokwenziwa kwe-AI kanye nezakhiwo

Ama-ejenti e-AI asusa isofthiwe kusuka kubasizi abangasebenzi aye ku- abahlanganyeli abazimele abangabona indawo yabo, bacabange ngemigomo eyinkimbinkimbi futhi bathathe izinyathelo egameni lethu. Kwabathuthukisi, lolu shintsho lushintsha konke: esikhundleni sokuxhumanisa imisebenzi eqhubekayo ezungeze i-LLM, uklama izinhlelo lapho imodeli ngokwayo iqhuba khona ukugeleza kokulawula, ihlela amathuluzi, futhi ibambisana namanye amanxusa nezinsizakalo.

Uma ufuna ukwakha okungathi sína, izinhlelo ze-ejenti zezinga lokukhiqiza, ukuqonda izinqubo ezintsha akusadingekiIzindlela ezijwayelekile zama-ejenti zokufinyelela amathuluzi (i-MCP), ukukhulumisana (i-A2A) nokuxhumana newebhu ngolimi lwemvelo (i-NLWeb) zisheshe zibe umgogodla "we-ecosystem yama-ejenti". Ngesikhathi esifanayo, usadinga ukwazi kahle izakhi eziyinhloko zama-ejenti ngokwawo: amamodeli, amathuluzi, imiyalelo, amaphethini okuqondisa kanye nezindlela zokuvikela.

Iyini ngempela i-ejenti ye-AI futhi ihluke kanjani ku-LLM ecacile?

I-ejenti ye-AI iqondwa kangcono njengohlelo oluphelele olwakhiwe nge-LLM, hhayi nje imodeli uqoboIncazelo eyamukelekile ngokwezemfundo (isibonelo kuStanford CS221) ichaza i-ejenti njengenhlangano yokubala etholakala endaweni, ekwazi ukuyibona ngezinzwa futhi isebenze kuyo ngama-actuator ukuze yandise amathuba empumelelo maqondana nomgomo othile.

Ngokwesofthiwe esebenzayo, ama-ejenti e-AI anamuhla ahlanganisa izithako ezine: a imodeli yolimi olukhulu ngokucabanga, ukufinyelela kumathuluzi angaphandle nama-API, uhlobo oluthile lwenkumbulo yokulandelela umongo ngokuhamba kwesikhathi, kanye nomgomo noma indima echazwe ngokucacile. Ngokungafani ne-chatbot elula ephendula imibuzo nje, i-ejenti ingahlela, ishayele amathuluzi, isabele emiphumeleni yawo futhi iqhube umsebenzi ngokuqhubekayo kuze kube yilapho kufinyelelwa umgomo.

Umthombo ovamile wokudideka ukuxuba “imodeli” kanye “ne-ejenti”. Imodeli efana ne-GPT‑4 noma i-Llama 3 "ingubuchopho" obunamandla kodwa obungenamsebenzi: ayenzi lutho uze uyithumelele isaziso, futhi ayikwazi ngokwayo ukuthumela ama-imeyili, ukushaya ama-API noma ukubuyekeza izizindalwazi. Ngakolunye uhlangothi, i-ejenti ihlanganisa imodeli ngendlela yokubona, ukucabanga kanye nesenzo. Isebenzisa izibikezelo zemodeli ukukhetha ukuthi yiliphi ithuluzi okufanele ilibize, ukuthi licele nini umsebenzisi ukuthi acacise, nokuthi liyeke nini.

Umehluko oyinhloko ukuthi ubani olawula umsebenziKusofthiwe yakudala, ikhodi yakho inquma ukulandelana: uma u-A bese kuba u-B bese kuba u-C. Ku-ejenti, i-LLM inquma ukuthi isinyathelo esilandelayo kufanele sibe yini ngokusekelwe esimweni samanje. Ingase ikhethe ukubheka i-oda, ukuvula ithikithi lokusekela, noma ukudlulisela icala komunye i-ejenti, konke lokhu kusuka esicelweni esifanayo sezinga eliphezulu.

Ama-ejenti nawo ayahlukahluka ngobunyoninco, kusukela ezinhlelweni ezilula zokusabela kuya ekufundeni, izakhiwo eziqhutshwa yimigomo. Indlela ejwayelekile yokulinganisa evela kuRussell noNorvig isasebenza ekuqondeni isimo sendawo: uthola ama-agent alula asabelayo (imithetho emsulwa uma-ke), ama-agent asabelayo asekelwe kumodeli (anesimo sangaphakathi esincane), ama-agent asekelwe emgomweni (ahlela umphumela ofiselekayo), ama-agent asekelwe kuzinsiza (athuthukisa amaphuzu ezinombolo ngaphezu kwemiphumela eminingi engenzeka) kanye nama-agent okufunda (avumelanisa inqubomgomo yawo ngokusekelwe kumpendulo).

Kungani izinqubo ezibalulekile enkathini yama-ejenti e-AI

Njengoba ama-ejenti eba namandla futhi esabalala, izinkinga ezintathu zivele ngokushesha: izindleko zokuhlanganisa, ukusebenzisana kanye nokuphepha. Ikhodi yokunamathisela engavamile yayo yonke i-API noma uhlelo lozakwethu ayilingani. Amafomethi obunikazi, asebenza kanye kuphela avimba ukubambisana phakathi kwamathuluzi nama-ejenti avela kubathengisi abahlukene. Futhi ukuhlanganiswa okusha ngakunye kwandisa ukuphepha kwakho.

Izinqubo ezigxile kuma-ejenti zihlose ukuxazulula ngqo la maphuzu abuhlungu ngokuchaza amazinga avulekile okuthi: indlela ababungazi abaveza ngayo amathuluzi kanye nomongo kuma-LLM (i-Model Context Protocol, noma i-MCP), indlela ababungazi abakhuluma ngayo nabanye ababungazi ngaphesheya kwemingcele yenhlangano kanye neyobuchwepheshe (i-Agent-to-Agent, noma i-A2A), kanye nendlela amawebhusayithi aveza ngayo okuqukethwe kwawo kanye nezenzo zawo ngendlela yolimi lwemvelo kuqala kubantu kanye nababungazi (i-Natural Language Web, noma i-NLWeb).

Kwabathuthukisi, lezi zinqubo zisebenza “njengama-adapter ajwayelekile” kanye “namakhadi ebhizinisi” amanxusa nezinsizakaloEsikhundleni sokufaka ikhodi yokuhlanganisa okuningi, uhlanganisa kanye namaseva e-MCP, ontanga abahambisana ne-A2A noma amasayithi e-NLWeb, bese uvumela iphrothokholi ukuthi iphathe ukutholakala, amakhono kanye nokuqinisekiswa. Lokhu kunciphisa kakhulu i-logic yokuhlanganisa ngokwezifiso futhi kukuvumela ukuthi ushintshe amamodeli noma amathuluzi ngaphandle kokubhala kabusha wonke amapayipi.

Ngesikhathi esifanayo, ukuphepha kwezinga lephrothokholi kuba yinto ebalulekileUkulawulwa kokufinyelela, ukuqinisekiswa okujwayelekile kanye nezincazelo zamandla ezicacile ku-protocol layer kwenza kube lula kakhulu ukucabanga ngokuthi ubani ongenza ini, kusuka kuphi, futhi ngaphansi kwaziphi izithiyo—okubalulekile ezilungiselelweni zebhizinisi lapho ama-ejenti angavunyelwa ukuthinta isitokwe, izinkokhelo noma idatha yamakhasimende ebucayi.

I-Model Context Protocol (MCP): i-adaptha yendawo yonke yamathuluzi nedatha

I-Model Context Protocol iyindinganiso evulekile echaza ukuthi izinhlelo zokusebenza zinganikeza kanjani amathuluzi kanye nedatha yomongo kuma-ejenti asekelwe ku-LLM. Ngomqondo, i-MCP ihlala phakathi kwama-ejenti akho nezinhlelo zakho ezikhona—izizindalwazi, ama-SaaS API, izinsizakalo zangaphakathi—futhi iziguqula zibe isethi yamakhono ahlanganisiwe, atholakalayo.

I-MCP ilandela ukwakheka kweklayenti neseva enezindima ezintathu eziyinhloko: i-host (uhlelo lokusebenza lwe-LLM olufana ne-IDE, iklayenti lengxoxo noma isikhathi sokusebenza se-ejenti) eqala ukuxhumana, izingxenye zeklayenti ezingaphakathi kwalowo msingathi ezigcina ukuxhumana komuntu nomuntu kumaseva e-MCP, kanye namaseva ngokwawo, okuyizinhlelo ezilula eziveza amakhono athile.

Ngaphakathi kwe-MCP, amaseva akhangisa izinto ezintathu eziyinhloko ama-ejenti angawasebenzisa ngendlela ehambisanayo: amathuluzi, izinsiza kanye nemiyalelo. Amathuluzi ayizenzo ezihlukile—“get_weather”, “purchase_product”, “search_flights”—anamagama, izincazelo kanye nama-schema okufaka/okukhiphayo. Izinsiza ziyizinto zedatha ezifundwayo kuphela njengamafayela, imigqa yedathabheyisi, noma amalogi, angaba umbhalo noma abe yi-binary. Imiyalelo iyizifanekiso ezichazwe ngaphambilini ezihlanganisa amaphethini obunjiniyela be-prompt noma ukugeleza kwezinyathelo eziningi.

Ukutholwa kwamathuluzi anamandla kungenye yezinto ezinkulu ezitholwe yi-MCPEsikhundleni sokubhala ikhodi eqinile yokuthi umsizi wokuhamba unomsebenzi we-“searchFlights” onesignesha ethile, i-ejenti ixhuma kuseva ye-MCP yenkampani yezindiza bese icela uhlu lwamakhono ayo. Iseva ibuyisela izincazelo zamathuluzi ezifundeka ngomshini, izimpikiswano zawo kanye nezimpendulo ezilindelekile. Lapho inkampani yezindiza ingeza ithuluzi elithi “upgrade_booking”, i-ejenti yakho iyalithola ngaphandle kokushintsha ikhodi, inqobo nje uma uhlonipha inkontileka ye-MCP.

I-MCP nayo ikholelwa ngamabomu ukuthi ayikholelwa kulokho okushiwoyoNgenxa yokuthi le phrothokholi imayelana namakhono kanye nomongo, hhayi mayelana ne-API yomthengisi oyedwa, iseva efanayo ye-MCP ingasetshenziswa kusuka kuma-LLM ahlukene noma ohlaka lwe-ejenti. Lokhu kukuvumela ukuthi uzame ukushintshana kwamamodeli noma amasu amamodeli amaningi (isb. ukusebenzisa imodeli encane, eshibhile yokugeleza okulula kanye nenamandla okucabanga okuyinkimbinkimbi) ngaphandle kokushintsha ukuhlanganiswa kwakho.

Enye inzuzo ukuphepha okujwayelekile. I-MCP ingafaka izindlela zokuqinisekisa ezihambisanayo, ezingagcinwa kangcono kunokuhlanganisa i-zoo yokugeleza kogunyazo okwenziwe ngokwezifiso kwe-API ngayinye yenkampani yangaphandle. Kumabhizinisi, lokhu kusho ukukala okuhlanzekile kusukela "ekuhlanganisweni okukodwa ekuhleleni" kuya "kumaseva e-MCP angamakhulu ekukhiqizweni" ngaphandle kokulahlekelwa ukulawula okhiye nezimvume.

Isibonelo esicacile senza indima ye-MCP icace: cabanga ngomsebenzisi ecela umsizi wokuhamba we-AI ukuthi “angitholele indiza esuka ePortland eya eHonolulu bese eyibhukha”. Umsizi, osebenza njengeklayenti le-MCP, uxhuma kuseva ye-MCP yenkampani yezindiza, abhale amathuluzi anjenge-“search_flights” kanye ne-“book_flight”, abize “search_flights” ngamapharamitha afanele, athole imiphumela ye-JSON, ayinikeze umsebenzisi, bese ebiza “book_flight” ngokusekelwe kunketho ekhethiwe. Umsizi akalokothi abize ama-API angaphakathi enkampani yezindiza ngqo; umane ukhuluma i-MCP.

I-Agent-to-Agent (A2A): inqubo yokusebenzisana kwama-ejenti amaningi

Nakuba i-MCP igxile ekuxhumeni ama-ejenti kumathuluzi nedatha, iphrothokholi ye-Agent-to-Agent imayelana nokuxhumanisa ama-ejenti komunye nomunye.. Lapho nje udlula i-monolithic "super-agent" ungene ku- uhlelo lwezinhlelo zokusebenza zama-ejenti akhethekile (ukuhamba, ukukhokhisa, ukuthutha, ukwesekwa…), udinga indlela ehlanzekile yokuthi batholane, bashintshisane ngomongo futhi babambisane emisebenzini ehlanganyelwe.

I-A2A iklanyelwe ukusekela lolu hlobo lokuhlelwa kwezinhlangano ezisabalalisiwe, ezihlanganisa wonke umuntu. Ivumela ama-ejenti avela ezinkampanini ezahlukene, ezindaweni zokubeka kanye nezindawo zokusingatha ukuthi asebenzisane ngesicelo somsebenzisi ngaphandle kokusebenzisa yonke indlela yokusebenzisana kusengaphambili. “I-Travel Agent” ehambisana ne-A2A ingabiza “I-Airline Agent”, “I-Hotel Agent” kanye “Ne-Car Rental Agent” eyakhiwe amaqembu ahlukene ngokuphelele.

Wonke umenzeli we-A2A udalula ikhadi le-Agent elifundeka ngomshini lokho kudlala indima efana nohlu lwamandla e-MCP, kodwa ezingeni le-ejenti hhayi ezingeni lethuluzi. Ikhadi le-Ejenti liqukethe igama le-ejenti, incazelo yolimi lwemvelo yalokho elikuphathayo, uhlu lwamakhono anezincazelo zokuthi kufanele libizwe nini, i-URL yalo yamanje, ulwazi lwenguqulo kanye namafulegi njengokuthi lisekela izimpendulo zokusakaza noma izaziso zokucindezela.

Ngasohlangothini lomuntu ofonayo, uMphathi Wezokuphatha unesibopho sokudlulisa umongo nokuphatha ukusebenzisana.Uma i-ejenti yendawo inquma ukuphathisa umsebenzi ongaphansi, umphathi wayo upakisha ingxoxo yamanje, isimo esifanele kanye nanoma yimiphi imikhawulo, bese eyithumela ku-ejenti ekude nge-A2A. I-ejenti ekude isebenzisa amathuluzi ayo angaphakathi kanye ne-LLM loop, bese ibuyisela umphumela ngaphandle kokuthi umuntu oshayelayo azi okungaphakathi kwayo.

Umphumela womsebenzi oqediwe kude ubuyiselwa njengento yobuciko. I-artefact ivame ukuhlanganisa umphumela womsebenzi, incazelo emfushane yalokho okwenziwe, kanye nomongo wombhalo ogeleze kuphrothokholi. Uma i-artefact isilethiwe, uxhumano lwe-A2A lungavala, lugcine ukusebenzisana ngakunye kusezingeni eliphezulu futhi kushibhile ngenkathi lusavumela ukubambisana okucebile.

Emisebenzini eqhubeka isikhathi eside noma engahambisani, i-A2A ivame ukuthembela emgqeni wemicimbiEsikhundleni sokugcina ukuxhumana kuvuliwe imizuzu embalwa ngenkathi i-ejenti ekude iqoqa idatha noma ilinde izinhlelo zangaphandle, umugqa womcimbi uphatha ukudluliselwa kwemiyalezo kanye nezibuyekezo. Lokhu kubaluleke kakhulu ezinhlelweni zama-ejenti amaningi ezisezingeni lokukhiqiza lapho ukuqina kwenethiwekhi, ukuzama kabusha kanye nokucindezeleka kwangemuva kubalulekile.

Izinzuzo ze-A2A zifana neze-MCP kodwa ezingeni le-ecosystemUthola ukubambisana okuthuthukisiwe phakathi kwama-ejenti angafani, ukuguquguquka kokukhetha i-LLM engcono kakhulu noma isu lokulungisa kahle i-ejenti ngayinye, kanye nokuqinisekiswa okwakhelwe ngaphakathi ukuze izingcingo zama-ejenti ahlukene ziphephe futhi zihlolwe. Kuba yinto enengqondo ukwakha "amaqembu ama-ejenti" ahlanganisa abathengisi abaningi kunokuzama ukufaka wonke amandla ku-monolith eyodwa.

I-Natural Language Web (NLWeb): okwenza kube lula ukusebenzisa ama-ejenti e-web

Iwebhu yakhelwe phezu kwamadokhumenti kanye ne-HTML, hhayi phezu kwezingxoxo kanye nama-ejentiAbasebenzisi bebelokhu bezulazula kumamenyu nasemabhokisini okusesha isikhathi eside ukuze bakhiphe ulwazi kumawebhusayithi, kuyilapho ukufinyelela okuzenzakalelayo kuvame ukuthembela ekukhucululeni okubuthakathaka noma kuma-API enziwe ngokwezifiso. I-NLWeb iphakamisa imodeli ehlukile: amawebhusayithi akhuluma ulimi lwemvelo, kokubili kubantu kanye nama-ejenti e-AI.

Ukufakwa kwe-NLWeb kuzungeze uhlelo lokusebenza oluphakathi lwe-NLWeb—ikhodi yesevisi eyinhloko ethola imibuzo yolimi lwemvelo, ixhuma kwisitoreji namamodeli, futhi ibuyise izimpendulo ezihlelekile. Ungacabanga ngayo njenge-"injini yolimi" yesayithi lakho, ukuhlela ukushumeka, ukusesha kwevektha kanye nokucabanga kwe-LLM.

Iphrothokholi ye-NLWeb ngokwayo ichaza imithetho eyisisekelo yalokhu kuxhumana kolimi lwemvelo. Ibeka ngendlela ejwayelekile indlela imibuzo ethunyelwa ngayo nokuthi izimpendulo zibuya kanjani, ngokuvamile njengefomethi ye-JSON kusetshenziswa amagama afana ne-Schema.org. Ngendlela efanayo yokwabelana ngamadokhumenti okujwayelekile kwe-HTML, i-NLWeb ihlose ukwenza ngendlela ejwayelekile ukufinyelela kokuqukethwe kwesayithi nezenzo okuqhutshwa ulimi, ivule indlela ye-"AI web".

Zonke izibonelo ze-NLWeb zisebenza njengeseva ye-MCPLokho kusho ukuthi ingaveza amathuluzi (njengendlela "yokubuza") kanye nezinsizakusebenza zedatha ezinhlelweni zangaphandle ze-AI nge-MCP. Ngokombono we-ejenti, isayithi lakho liba nje elinye iphuzu lokugcina le-MCP: lingabiza "ukubuza" ngombuzo, lithole impendulo ehlelekile ehlobene nokufakwa kwangempela kukhathalogi yakho, futhi ligweme ukuphupha ngemikhiqizo noma amakhasi angekho.

Ngaphansi kwe-hood, i-NLWeb incike kakhulu kumamodeli okushumeka kanye nedathabheyisi ye-vectorUma udla okuqukethwe kwesayithi lakho—uhlu lwemikhiqizo, izincazelo zehhotela, okuthunyelwe kwebhulogi—i-NLWeb ikuguqula kube ukushumeka kwe-vector futhi ikugcine esitolo se-vector esihambisanayo njenge-Qdrant, i-Milvus, i-Azure AI Search, i-Snowflake noma i-Elasticsearch. Ngesikhathi sokubuza, ithola izinto ezifanayo kakhulu bese izidlulisela, kanye nombuzo womsebenzisi, ku-LLM ukuze yenze impendulo esekelwe kokuqukethwe kwangempela.

Isayithi lokubhukha uhambo liyisibonelo esihle se-NLWeb esebenza kahle. Ufaka idatha ehlelekile yezindiza, amahhotela namaphakheji (ngokusebenzisa i-Schema.org noma ama-RSS feeds), udale ukushumeka bese uwagcina. Uma umsebenzisi ethayipha ukuthi “ngitholele ihhotela elinobungani nomndeni e-Honolulu elinechibi lokugeza ngesonto elizayo” ebhokisini lengxoxo, i-NLWeb ibuza isitolo se-vector samahhotela afanele, ivumele i-LLM ihumushe “elinobungani nomndeni” kanye neminye imikhawulo ethambile, bese ibuyisela impendulo yolimi lwemvelo esekelwa yimpahla yangempela. Isibonelo esifanayo se-NLWeb, ngokusebenzisa isikhombimsebenzisi sayo se-MCP, sivumela i-ejenti yokuhamba yangaphandle ukuthi ibuze, isibonelo, mayelana nezindawo zokudlela ze-vegan eduze kwalawo mahhotela futhi ithole i-JSON eqhubekayo, esetshenziswa ngomshini.

Uma kunengqondo ukwakha i-ejenti ye-AI nhlobo

Akuzona zonke izinkinga ezidinga i-ejenti; ngezinye izikhathi isevisi elula yokunquma ingconoAma-ejenti ayakhanya lapho umsebenzi ungabanjwa kalula njengeqoqo lemithetho eqinile, lapho kuncike kakhulu kudatha engahlelekile, noma lapho inani lezinto ezihlukile kanye namacala onqenqema kwenza ukulungiswa kwemithetho kube buhlungu.

Imindeni emithathu yamacala okusetshenziswa ifaneleka kakhulu kuma-ejenti: ukwenza izinqumo eziyinkimbinkimbi (isibonelo, ukunquma ukuthi kufanele kuvunyelwe yini ukubuyiselwa kwemali kwamakhasimende ngaphansi kwezinqubomgomo ezicacile), imithetho enzima ukuyigcina (njengokubuyekezwa okuyinkimbinkimbi kokuphepha kwabathengisi noma ukuhlolwa kokuthobela imithetho), kanye nokugeleza okubuswa ulimi lwemvelo (ukucubungula izimangalo, izicelo zamakhasimende ezikhululekile, imisebenzi yocwaningo).

Indlela ewusizo yokuguqula izinto ibe yi-heuristic ukubheka izinhlelo ezikhule ngama-patches angapheli kanye nemithetho ekhethekileUma ngisho nonjiniyela abaphezulu behluleka ukubikezela ukuziphatha noma ukufaka ikhodi yezinguquko ezintsha zenqubomgomo ngaphandle kokuphula okunye, kungenzeka ukuthi inkinga eyisisekelo iwukuchaza amagama, hhayi okunengqondo kuphela. Leyo yindawo ephelele ye-ejenti eqhutshwa yi-LLM engacabanga ngombhalo, izinqubomgomo nezibonelo.

Ngokuphambene nalokho, emisebenzini eqinisekile kakhulu enokufakwayo okucacile kanye nemiphumela, ikhodi yakudala ngokuvamile izoba eshibhile, esheshayo futhi ethembekile kakhuluUma umsebenzi wakho “ukuguqula le nombolo ibe yifomethi ehlukile” noma “ukusebenzisa lo mbuzo we-SQL bese ubuyisela imigqa”, ukwengeza iluphu ye-ejenti phezulu cishe kuyinkimbinkimbi engadingekile.

Izakhi eziyinhloko ze-ejenti ye-AI

Naphezu kokuduma okukhulu, isakhiwo sangaphakathi se-ejenti eklanywe kahle silula impelaCishe wonke amaphethini ancipha abe yizinsika ezintathu: imodeli eyenza ukucabanga, amathuluzi axhumanisa nomhlaba wangaphandle, kanye nemiyalelo evimbela futhi eqondisa ukuziphatha.

Imodeli iyinjini yokwenza izinqumoAma-LLM ahlukene ashintsha ikhwalithi yokucabanga, ukubambezeleka kanye nezindleko. Isu elivamile nelisebenzayo yileli: qala ngemodeli enekhono eliphezulu lokusungula isisekelo sekhwalithi bese uqonda ukuthi "okuhle" kubukeka kanjani esizindeni sakho, bese uhlola kancane kancane amamodeli amancane noma ashibhile emisebenzini engaphansi njengokuhlukanisa noma ukubuyisa lapho kungadingeki khona ukucabanga okuphezulu.

Amathuluzi andisa umenzeli ngale kombhalo omsulwa. Yimisebenzi, ama-API, noma izinsizakalo i-ejenti engazishayela ucingo: ukubuza isizindalwazi, ukuthumela i-imeyili, ukusesha iwebhu, ukusebenzisana ne-UI yakudala ngemodeli yokusebenzisa ikhompyutha, njalo njalo. Amathuluzi aklanywe kahle aqoshwe phansi, angasetshenziswa kabusha kuwo wonke ama-ejenti futhi abonakala kahle ngamaphrothokholi ajwayelekile njenge-MCP.

Imiyalelo iyingxenye enganakwa kakhulu ye-ejenti. Udinga okungaphezu kokuthi “ube wusizo”. Imiyalelo esezingeni eliphezulu ichaza indlela yokuhlukanisa imisebenzi, indlela yokuziphatha lapho ulwazi lungekho, ukuthi yimaphi amathuluzi okufanele uwakhethe ezimweni ezinjani, ukuthi yini ebalwa njengempumelelo, nokuthi yini okufanele uyigweme. Amaqembu amaningi asebenzisa ngempumelelo ama-SOP akhona, amadokhumenti esikhungo sosizo noma izincwadi zokudlala zangaphakathi ngokuziguqula zibe iziqondiso ezinobungani ne-LLM, ezibhalwe izinombolo imodeli engazilandela.

Sekuvamile kakhulu ukukhiqiza noma ukulungisa imiyalelo ngokuzenzakalela kusetshenziswa ama-LLM ngokwawoIsibonelo, ungafaka isihloko sesikhungo sosizo ku-meta-prompt ecela imodeli ukuthi iyibhale kabusha njengesethi ecacile, enezinombolo yemiyalelo ye-ejenti, okuhlanganisa nokuphathwa okucacile kwamacala e-edge. Lokhu kugcina ukuziphatha kuhambisana nedokhumenti yakho njengoba ithuthuka.

Amaphethini okuhlanganisa: izinhlelo ze-single-ejenti vs multi-agent

Ngaphansi kwe-hood, ama-ejenti asebenza ngendlela ejikelezayo: qaphela isimo samanje, nquma ukuthi wenzeni, yenza okuthile (ngokuvamile ngethuluzi), buyekeza umongo, bese uphinda kuze kube yilapho kufinyelelwa isimo sokuma (umgomo ofinyelelwe, iphutha, ukungenelela komsebenzisi, noma uhambo lokuqapha). Lokhu "ku-agent loop" yilokho okuguqula ucingo lwe-LLM oluhamba kanye lube injini yokusebenza eqhubekayo.

Ukwakhiwa okulula kakhulu yi-ejenti eyodwa enamathuluzi. Ithola imiyalezo yomsebenzisi, izizathu ngayo, inquma ukuthi imaphi amathuluzi okufanele iwashayele, futhi ibuyisele izimpendulo. Amafreyimu avame ukudalula ingxenye yomgijimi eqhubeka ishayela imodeli kuze kube yilapho imfuneko ethile yokuqeda igcwalisekile—njengokuthi “akusekho ukubiza kwamathuluzi awusizo” noma “umphumela ohleliwe okhiqizwe”. Le phethini ilungele izinguqulo zokuqala kanye nezinkinga ezibanzi.

Njengoba ubunzima bukhula, amaqembu avame ukuthuthela kuma-topology ama-ejenti amaningiKunezinhlobo ezimbili eziyinhloko. Ngokwephethini yomphathi, i-ejenti “ye-orchestrator” ephakathi idlulisela imisebenzi emincane kuma-ejenti akhethekile abonakala njengamathuluzi—ake sithi, abahumushi ezilimini ezahlukene, i-ejenti yocwaningo kanye nomgxeki. Umphathi ugcina ukulawula komhlaba wonke futhi uhlanganisa konke ndawonye.

Iphethini yesibili ihlukaniswe kakhulu. Lapha, ama-ejenti anikeza ontanga umsebenzi uma bebona ukuthi isicelo singaphandle kwesizinda sabo. I-ejenti yokuhlunga ingathumela imiyalezo yamakhasimende kuma-ejenti okusekela ubuchwepheshe, okuthengisa noma okuphatha ama-oda, ngayinye inemiyalelo namathuluzi ayo. Ukuhamba kokulawula kuyagxuma phakathi kwama-ejenti ngaphandle komhleli oyedwa ophakathi.

Womabili amaphethini angahlangana ngokwemvelo ne-A2A ngezinga elikhuluNgaphakathi komkhawulo womkhiqizo noma wesevisi encane ungasebenzisa imodeli ye-orchestrator-plus-specialists, kanti kuzo zonke izinkampani noma iminyango uthembele ku-A2A ukuze ukhulume nama-ejenti angaphandle akhangisa amakhono awo ngama-Agent Cards.

Izithiyo zokuvikela: ukugcina ama-ejenti azimele ephephile futhi ethembekile

Ukunikeza amanxusa ukuzimela kusho nokwamukela izingozi ezintsha: bangase baveze idatha ebucayi, benze izinguquko ezingagunyaziwe, noma bathathe izinyathelo ezithinta ezezimali noma idumela. Izithiyo zokuvikela ziyisivikelo esilawula lezi zingozi ngaphandle kokunciphisa ukusebenza kwe-ejenti.

Umklamo wokuzivikela uvame ukuhilela izendlalelo eziningi zokuvikelaAbanye basebenza kokufakwayo (ukuvimba noma ukuhlanza izicelo ezinonya noma ezingaphandle kobubanzi), abanye ezinqumweni zemodeli eziphakathi nendawo (ukuhlola ukuthi isenzo sivunyelwe yini ngaphambi kokusenza), kanti abanye ezikhishweni (ukuhlunga ukuphepha, ukuthobela imithetho noma ukuvuza kwedatha ngaphambi kokuba izimpendulo ziphume ohlelweni).

Ezintweni eziningi ezisetshenziswayo, izivikelo zisebenza “ngokuhambisana” nentuthuko enethemba ye-ejenti. Iluphu ye-ejenti iya phambili, kodwa izinyathelo ezithile—njengokubiza kwamathuluzi okungase kuhlele idatha—zihlanganiswe nokuhlolwa kwe-guardrail. Uma i-guardrail ithola ukwephulwa, ingamisa isenzo, iphakamise okuhlukile, noma ikhuphukele ku-opharetha ongumuntu.

Ezinye izivikelo zokuvikela ngokwazo ziqhutshwa yi-LLM ezigxile kuzo imikhawulo kanye nezingozi noma ngisho nama-ejentiIsibonelo, ungase ugcine i-ejenti yokutholwa kwe-churn ezinikele ehlola imiyalezo yamakhasimende engenayo futhi iveze leyo ekhombisa ingozi ephezulu yokukhanselwa. I-guardrail esezingeni eliphezulu bese isebenzisa lesi siginali ukuqala imisebenzi yokugcina noma idinga ukubuyekezwa okuphoqelekile komuntu ngaphambi kokuvala ukusebenzisana.

Izithiyo zokuvikela ezisebenzayo zihlanganisa nemingcele eqinile kanye nezivalo zokuphunyuka. Izinyathelo eziphezulu zibalwa ukuze kugwenywe izihibe ezingenamkhawulo, imikhawulo esekelwe engcupheni ephoqa ukuvunyelwa kwabantu ngezenzo ezibucayi, kanye nokubuyela emuva okucacile lapho ukuzethemba kwemodeli kuphansi konke kunegalelo ekusetshenzisweni okuphephile ezindaweni zangempela.

Kusukela kumbono kuya ekusebenzeni: umklamo ohamba kancane we-ejenti yokusekela i-oda

Ukuze usekele le mibono, cabanga ngokuvela kohlelo lokusekela ama-oda esitolo esiku-inthanethiInguqulo yokuqala ivame ukuba yindawo yokuphela esabelayo: uma unikezwe i-ID ye-oda, thatha isimo sayo kusizindalwazi bese uyibuyisela. Akukho ukucabanga, akukho nkumbulo, futhi akukho ukuhamba komsebenzi—lokhu akukabi yi-ejenti.

Isinyathelo sokuqala se-ejensi ukuvumela imodeli ukuthi ilawule ukuhamba komsebenziEsikhundleni sokucabanga ukuthi i-ID ye-oda ikhona, unikeza imodeli ingxoxo ephelele bese uyivumela ukuthi inqume ukuthi yenzeni. Uma umsebenzisi ebuza ukuthi “Liphi iphakheji lami?” ngaphandle kokunikeza i-ID, imodeli ingakhetha isenzo esithi “ASK_FOR_ORDER_ID” bese icela umsebenzisi ukuthi athole ulwazi olwengeziwe.

Okulandelayo, ugoqa lokhu kucabanga ngendlela ejikelezayo bese wethula isimoNgemva komyalezo ngamunye womsebenzisi noma ucingo lwethuluzi, i-ejenti iphinda ihlole isimo. Ingase ilande i-oda, ibuyekeze umongo, ihlole ukuthi inolwazi olwanele yini lokuphendula, noma ibuze umbuzo olandelayo. I-loop iyama kuphela uma impendulo ecacile isithunyelwe noma kufinyelelwe umbandela wokuqeda.

Njengoba ububanzi bukhula ngale kokuhlolwa kwesimo, i-ejenti iqala ukukhetha amathuluzi ngokususelwa kunhloso. Inkinga yokuthumela ingase idluliselwe ku-“open_incident”, isicelo sokubuyiselwa imali ku-“initiate_refund”, kanye nombuzo wesimo olula ku-“get_order_status”. Awufaki ikhodi kusihlahla esinqunyiwe samagatsha e-if-else; kunalokho, imodeli ikhetha izenzo kumenyu yamathuluzi achazwe nguwe noma atholwe nge-MCP.

Kuleli qophelo wethula izivikelo zokuvikela kanye nokuhlolwa kwengozi ezungeze amathuluzi abucayi. Imisebenzi yokufunda kuphela ingase yenziwe ngqo, kodwa noma yini eshintsha isimo (ukukhipha imbuyiselo, ukukhansela ama-oda, ukuguqula amakheli) idlula endaweni yokuqapha eqaphela ubungozi. Izenzo ezinobungozi obukhulu zidinga ukuvunyelwa ngabantu; izenzo ezinobungozi obuphakathi zingase zibangele ukuqinisekiswa okwengeziwe; izenzo ezinobungozi obuphansi zingaqhubeka ngokuzenzakalelayo.

Ekugcineni, ubeka imingcele yokusebenza kanye nemithetho yokudluliselana kwabantuUma umenzeli ethola inani elikhulu lemizamo ehlulekile, ehlangabezana nolwazi oluphikisanayo, noma ebhekene nesinqumo esiyingozi kakhulu ngaphandle kwesibopho sakhe, udlulisela kumenzeli wokusekela abantu onawo wonke umongo oqoqiwe. Le ndlela ehlanganisiwe ikuvumela ukuthi usebenzise ngokuphephile ukuzimela ngenkathi ugcina ukulawula amacala onqenqema.

Izinhlaka zokucabanga ezithuthukisiwe kanye namathuluzi e-ejenti yesimanje

Ngaphezu kwalezi zisekelo zokwakha, izinhlaka zokucabanga ezithuthukisiwe zisiza ama-LLM ukuthi aziphathe njengezithunywa ezihleliwe kunezibikezelo zebhokisi elimnyamaAmaphethini amabili adumile yi-Chain-of-Thought (CoT) kanye ne-ReAct (Reason + Act).

I-Chain-of-Thought imane icele imodeli ukuthi icabange isinyathelo ngesinyathelo, ukuhlukanisa imibuzo eyinkimbinkimbi ibe yizinyathelo zokucabanga eziphakathi ngaphambi kokukhiqiza impendulo yokugcina. Ucwaningo lubonisa ukuthi lokhu kungathuthukisa kakhulu ukusebenza emisebenzini enzima yokucabanga kumamodeli amakhulu, futhi kuhambisana ngokwemvelo ne-agent loop: ucingo ngalunye lwamathuluzi luhambisana nochungechunge olubanzi lokucabanga.

Phendula ngokuqinile imibhangqwana icabanga ngokusebenzisa amathuluzi. I-ejenti ishintshashintsha ngokusobala phakathi kwemicabango, izenzo kanye nokubona: ichaza lokho ehlose ukukwenza, ibiza ithuluzi, ihlola umphumela, futhi ibuyekeze uhlelo lwayo. Le phethini isekela izinhlelo eziningi ze-ejenti ezizimele zakuqala njenge-AutoGPT kanye ne-BabyAGI, ezikhiqiza futhi zibeke phambili uhlu lwezinto okufanele zenziwe ukuze zifinyelele umgomo womsebenzisi.

Izinhlaka zesimanje nama-SDK zihlanganisa le mibono ibe yizifinyezo ezinobungani nonjiniyela. Amalabhulali afana ne-LangChain, i-LangGraph, i-CrewAI noma amathuluzi amancane esitayela se-“smolagents” ahlinzeka ngamabhlogo okwakha ukubiza amathuluzi, imisebenzi esekelwe kugrafu, ukuhlelwa kwama-ejenti amaningi kanye nenkumbulo eqhubekayo. Amaningi alawa ma-toolchain afaka nesiqondiso se- ama-ejenti ngokwezifiso ku-VS CodeAmapulatifomu obunikazi avela kubathengisi bamafu nabadlali abanjengo-OpenAI anezela izakhiwo ezisezingeni eliphezulu zama-ejenti, izithiyo zokuvikela kanye nokuhlola.

Okubaluleke kakhulu, lezi zinhlaka ziya ngokuya zihlangana nezinqubo ezifana ne-MCP, i-A2A kanye ne-NLWebEsikhundleni sokubhaka izixhumi ezisebenza kanye kuphela, ama-ejenti angaxhuma ezingqimbeni zamandla ezijwayelekile, akhulume nama-ejenti angaphandle ngokusebenzisa ama-Agent Cards, futhi aphathe amasayithi anikwe amandla yi-NLWeb njengama-API olimi lwemvelo, asezingeni eliphezulu. Lokhu kuhlangana phakathi kwamaphrothokholi kanye namathuluzi yikho okuvumela i-ecosystem yama-ejenti amakhulu, asebenzisanayo.

Konke lokhu kusekelwe ekuqhubekeni kusukela ekungabhalweni kwekhodi kuya ezixazululweni ezinekhodi ephezuluAmapulatifomu abonakalayo esikhaleni esingenakhodi avumela abangebona onjiniyela ukuthi babhale imisebenzi ye-ejenti namathuluzi anezixhumi zokudonsa nokulahla kanye nokulungiselelwa kolimi lwemvelo. Ngakolunye uhlangothi, izindawo ezinamakhodi aphezulu zinikeza onjiniyela ukulawula okunembile kokuhlelwa, ukuhlolwa kanye nokusetshenziswa, ngokuvamile kuhlanganisa izinhlaka nengqalasizinda eyenziwe ngokwezifiso ku-AWS, Azure noma amafu afanayo.

Kuyo yonke le mikhakha, izinhlangano eziwinayo yilezo ezifunda ukwakha ama-ejenti, hhayi nje ukuwasebenzisa.Ukuqonda amaphrothokholi, amaphethini kanye nezindlela zokuvikela kukuvumela ukuthi weqele ngale kokuhlola "ukuzama i-chatbot" futhi uye e-automation eqinile, engandiswa: kusukela kuma-ejenti okuhlaziya angaphakathi kanye nabashayeli bezindiza abangonjiniyela, kuze kube yizinhlelo zama-ejenti amaningi ezihlanganisa isitokwe, izinkokhelo kanye nokuhlangenwe nakho kwamakhasimende ngesikhathi sangempela. Njengoba ama-ejenti eqhubeka nokuvuthwa, lawa makhono okuklama aba yimpumelelo yangempela yokuncintisana.