Introduction Hey bonke, ndiye Max Nechaev - umlawuli wobuchwepheshe kwi-Snoonu, umlawuli, kunye ne-AI enthusiast. Ndingathanda kwi-development ye-software ngexesha elide. Emva kokufaka i-app ye-iOS, kukho amaxabiso ezininzi ze-architectural. Zifunyenwe ngokufanelekileyo. Thought through. Battle-tested ngexesha elide. Kwixesha elidlulileyo, ngexesha elidlulanga ukuba ndingathanda ukuba ndingathanda i-architecture entsha - ikakhulukazi akukho kwi-AI systems. Kodwa ngaphambi kokufikelela kwi-solution, nqakraza malunga ne-problem. Okwangoku, i-AI isebenzisa i-hype ye-cycle ye-crazy. Iimveliso ze-start-up ziyafumaneka kwaye zithunyelwa ngosuku zonke. Abantu abaninzi zithunyelwa kwi-low-code kunye nezixhobo ze-no-code ezifana neMake, n8n, kunye nabanye. Kwaye uyazi ukuba? Ndingathanda ukuba kuhle. Izigidi ezininzi zithunyelwa kwiintlobo ezintsha ze-reality. Kubonakala njengabaxesha elidlulileyo le-internet, xa iwebhu zangaphakathi kwi-static ukuya kwi-interactive, kwaye ngexesha elifutshane ndingathanda ukuba nayiphi na into kunokwenzeka. Izixabiso ezininzi ziquka i-AI agents. Kwaye ndiyifunayo. Kodwa kukho i-catch. Akukho kwakhona ukufumana indlela yokwakha iinkqubo ze-AI ezincinane, ezincinane. Akukho iimfuno ze-design. Akukho isakhiwo esisetyenziswa kakhulu. I-vibes kuphela. I-Agents eyenziwe kwi-n8n ibonakala njenge-spaghetti. I-impossible yokufunda, i-debug, okanye i-scale. I-Tutorials kwi-building agents kwi-Python? I-story efanayo. Yonke into efakwe kwifayile elinye, akukho ukuxhaswa kweengxaki. Ngokuqhelekileyo, i-developer ezijolileyo ingathanda ukuba ziyafumaneka izinto - kodwa abapheliyo akayi. Kwaye zithuba i-building iinkqubo ezininzi ezininzi. Ngoko namhlanje, ndiyabakhokela isisombululo ngathi ndiyabakhokela. Ngoko ndiyabakhokela iinkqubo zayo ze-AI. Ngoko ndiyabakhokelela iingxaki. Ngoko ndiyabakhokelela iingxaki, i-scalable, kwaye ngokwenene ngokwenene. Nceda uqhagamshelane nathi kwi-framework yam, i-design pattern, okanye i-architecture - uqhagamshelane nangokufuneka - Iingxowa zeengxowa zeengxowa. AAC Meet AAC (Agent Action Chains) Kwixesha elinye, ndingathanda into elula. Ukuba i-chaos ibonakala kwakhona kwaye kwakhona, ndingathanda ukuba ayikho i-accident. Ngoko ke kunokuba kunceda ukuba i-system ayikho. Ngoko ndiqala ukuhanjiswa. Okokuqala, kunye neeblokhi ngaphakathi n8n. Emva koko ndiqala ukuhanjiswa kwe-logic ngamanzi - apha ngoko ndiyabakhokela iingxaki, apha ndiyabakhokelela, apha i-logic core. Ekugqibeleni, iingxaki zangaphambili zangaphambili. Kwaye ngoko ke iye ndiza. Oku kufuneka isakhiwo. Oku kufuneka i-architecture. Yintoni i-AAC - i-Agent Action Chains. Yintoni iinkqubo ezininzi kunye neengxasi ezininzi ze-docs. I-AAC yindlela yokwakha izinhlelo ze-AI njenge-engineer ebandayo, ayikho njenge-omnye ebandayo izinto kunye nokufuna kwimveliso engcono. Yintoni i-idea core? Ndibandakanya i-system kwi-agents kunye neengcali ezininzi. Yonke i-agent inenza into elifanayo. Zonke zihlala nge-JSON contracts - akukho ukufikelela, akukho umnqweno, akukho umnqweno. Zonke into kubaluleke, ukuvelisa, kwaye kulula ukucacisa. Nceda ucinga nje njenge-team. Umgangatho umgangatho umgangatho Zonke iinkcukacha zokuvelisa. Umntu omnye uye wahlala into elandelayo. I-Agent ye-Memory isixazululo. Umnqweno umnqweno umnqweno. Umntu omnye ukhangela kwaye ukhangela yonke into. Wonke umntu uyazi umsebenzi wayo. Kwakhona, inkqubo isebenza njengoko iiyure yeSwiss. Umbala we-AAC yinto ukuba i-platform-agnostic. Ungasebenza kwi-n8n, kwi , ePython usebenzisa FastAPI, nangona ngaphakathi LangChain okanye LangGraph ukuba uthathe ixesha ukutshintsha ngokufanelekileyo. Yenza.com Here’s How It Works (In Plain English) Ukukhanyisa iqela eluhlaza. Akukho iqela eluhlaza apho wonke umntu usebenza yonke into ngexesha elinye, kodwa iqela apho wonke umntu uyazi umsebenzi wayo, unayo indawo yayo, kwaye inikeza imiphumo eluhlaza. Umntu uthatha imibuzo, umntu omnye uthetha, umntu omnye uthetha, umntu omnye uthathe i-double-checks, kwaye umntu uthatha i-log yeyona kwakhona. Yintoni nje i-AAC isebenza. Xa inguqulelo ifumaneka kwi-system - okanye kwi-user okanye i-service ye-external - kuqala kuza nge-agent enokufumana i-death kwaye i-off. Emva koko i-processing yokufakelwa. Thina akufutshane iziphumo. Okokuqala, thina ukwakha inguqulelo, ukunciphisa i-noise, ukuhlaziywa. Oku kuphela isinyathelo ukunciphisa iiyure ze-debugging emzimbeni. I-input epheleleyo i-half of the architecture. Okulandelayo i-orchestrator. I-brain ye-system. It is not executing the task itself. Ngokwenza oku, ibonelela ngubani kufuneka ukwenza into, ngexesha elizayo, kunye neengxaki. I-orchestrator's role is to plan, direct, and manage the flow. Nothing more, nothing less. I-orchestrator's role is to plan, direct, and manage the flow. Emva koko iingcali zihlanganisa. Lezi zengxaki zihlanganisa. Umntu uqhagamshelane, omnye uqhagamshelane, omnye uqhagamshelane iinkcukacha zeengxaki, njl. Wonke unayo isicelo esibonakalayo. Abanikezela, abanikezela, abanikezela, abanikezela, abanikezela, abanikezela, abanikezela, abanikezela, abanikezela, abanikezela, abanikezela, abanikezela, abanikezela, abanikezela, abanikezela, abanikezela, abanikezela. Nangona i-system kufuneka ibhalisele into, i-memory ibhalisele. Oku yi-agent yayo yayo. Akukho ibhalisele yonke into ngokufanelekileyo. Ngoku, uyakwazi ukufumana izinto ezifanelekileyo ngexesha elifanelekileyo - iingxowa ezidlulileyo, i-context ye-user, iimpawu ze-internal. Ngoko ke i-system yakho ibhalisele ukuba i-goldfish kunye nokufumana njenge-assistant efanelekileyo. Umgangatho we-error ibhaliswe ngokufanelekileyo. I-AAC ibandakanya i-agent eyenziwe ngoko - i-Guard. I-Monitoring for failures, catches exceptions, isebenza xa kukho into. It is not a patch, not a random try-catch. It is a real part of the architecture. And it is responsible for resilience. Nangona konke oku kutyelela, i-Observer ikhawulezela ngokucacileyo. I-Observer ikhawulezisa into yeenkcukacha, ixesha elide, izixazululo zenzelwe. Akukho kuphela ngenxa yokuzonwabisa - oku kukunika ukubonisa kwinkqubo. Akukho "ukusebenza?" kodwa Ukusebenza ukuba, Yiba, kwaye apho kunokufumana. Yintoni Yintoni Kwaye ekugqibeleni, xa inkampani ephelele umsebenzi yayo, umphumela uqhagamshelane kunye nokuthunyelwe - kwakhona kwi-user, kwi-API response, okanye apho kufuneka uqhagamshelane. Clean, eyakhelweyo, kwaye lula ukutya. Yintoni i-AAC. I-system eyenziwe ukuba isebenze nje, kodwa ukwandisa, ukwandisa, kwaye ibekwe phantsi kolawulo. All AAC Roles: Who Does What and Why It Matters I-AAC ibekwe kwi-separation clear of responsibilities. I-Agent ye-block kuphela ye-technical - i-unity ye-logic kunye ne-roll defined kunye ne-communication contract. Le nkqubo ibonise ukuba i-rolls ezininzi ziyafumaneka, indlela yokusebenzela, yaye ngoko ukuchithwa kwe-one of them ngokuvamile kuholele kwi-architectural pain emva. Ingress Agent: the Entry Point Ukulungiselela i-signal yokufikelela. Purpose: Yinto ingaba i-webhook, i-form submission, i-Telegram message, i-Kafka event - akukho nto. Yintoni oku kubaluleke i-isolating input handling ukusuka kwi-logic yebhizinisi. Ngenxa yokufaka i-logic kunye ne-integration layers yindlela yokuqala yokwakha i-monolith. Kwaye i-monoliths akufutshane-okufutshane. Kwi-n8n, oku kubandakanya kuphela i-webhook node. I-webhook ithatha i-request, i-logs yeedatha kunye nokuthumela i-powerload ye-JSON evulekileyo kwi-system. Umzekelo (pseudocode): { "source": "user", "payload": { "message": "I’d like to return my order" } } Preprocessing Agent: the Filter and Normalizer Yenza ingxubevange ingxubevange into efanelekileyo. Purpose: I-Agent yenza i-noise. I-Fixing i-casing, i-strips ze-symbols ezizodwa, i-validation ye-structure, kunye ne-extracts yeengxaki ze-key. I-Fixing i-cockpit ibonise ukuba zonke iingxaki ezivela kwi-system zibonakalayo kunye ne-consistent. Inani abaninzi bafumane i-preprocessing. I-great mistake. Akukho kuphela i-" nice-to-have" - nto leyo isiseko. Ngokutsho xa unxibelelana ne-LLMs, isifo se-context yinto yonke into. I-input eyodwa enxulumeneyo kunye nenkqubo epheleleyo ingangena. Kwi-n8n, oku ngokuvamile i-function node. Kwi-code, ingaba i-middleware, i-handler, okanye i-class ye-utility encinci. const cleanInput = input.message.trim().toLowerCase() Orchestrator Agent: the Brain of the System Qinisekisa ukuba into kufuneka kuba. Purpose: I-orchestrator ayenza umsebenzi wayo. Izixhobo zayo zihlanganisa, ukuhambisa kunye nokuhambisa. I-orchestrator ithatha i-input eyenziwe, i-intent ye-user, kwaye i-agent ye-specialist efanelekileyo. Kwiimeko ezincinane, oku kunokuba kuphela if/else block. Kwiimeko ezininzi ezihlangeneyo, i-LLM epheleleyo esebenza kunye neengxaki ezilungileyo. Umzekelo, unako ukufumana i-GPT-4 yokuguqulwa kwiphaneli ye-execution e-JSON: { "steps": [ {"agent": "Classifier", "input": {...}}, {"agent": "Memory", "input": {...}}, {"agent": "Responder", "input": {...}} ] } Uyakwazi ukufumana iindidi zeengcali ezisetyenziswa kunye nokukhetha indlela efanelekileyo. Umzekelo wokugqibela: I-Orchestrator ayisombulula inkcubeko. I-Orchestrator ibekwe i-route. I-Orchestrator ibekwe ngubani kufuneka ukwenza njani, kwaye ngexesha elizayo. Specialist Agents: Focused Experts (TOOLS) Yenza umsebenzi olufanelekileyo, olufanelekileyo. Purpose: Zonke iingxaki ze-Agents ziye ziyafumaneka ukuyisombulula yonke inkinga - nje ingxaki elinye elifanelekileyo, kwaye ziyafumaneka kakuhle. Ukulungiselela isicelo Yenza ingxoxo Izakhiwo ze-extraction Ukukhangela i-API ye-external Ukulungiselela umbhalo Ingaba unako ukufikelela kwiintlobo ezininzi zezi. Yonke i-function eyahlukileyo, i-module evulekileyo, okanye i-subflow kwi-n8n. Kwakhona yintoni oku kwenza inkqubo efanelekileyo kunye ne-scalable. Ingaba ufuna ukongeza iinkonzo ezintsha? Nceda uqhagamshele i-specialist entsha kwaye uqhagamshele. I-orchestrator ayikho kufuneka ifunyenwe - kufuneka nje ukwazi ukuba le mveliso ikhona. Umzekelo we-contract phakathi kwe-orchestrator kunye ne-specialist: { "agent": "Classifier", "input": { "text": "My order never arrived" }, "output_expected": { "label": "delivery_problem" } } Memory Agent: the Interface to the Past Yenza iinkcukacha ezifanelekileyo. Purpose: I-Agent yenza iimpawu. Yinto. I-Search ye-database, i-vector store, i-Redis cache - yintoni nawe - kwaye ibonelela ngexabiso efunyenweyo. Ayikwazi ukuthatha imibuzo, ayikwazanga, ayikwazanga. Khetha kuphela idatha yokwenene. Ingaba usebenzise ukufumana iintlawulo ezidlulileyo yabasebenzisi, ukuxhaswa kwe-vector embeddings kwi-knowledge base, okanye ukufumana iintlawulo ezidlulileyo. Kwi-AAC, i-memory iyiphi na ingxaki yayo. Ngoko ke, unako ukucacisa, ukucacisa, okanye ungenza kwiinkqubo ezininzi. SELECT * FROM orders WHERE user_id = "user_42" ORDER BY created_at DESC LIMIT 3 Guard Agent: Protection from Chaos Ukulungiselela kunye nokufumana iingxaki. Purpose: Ukuba i-specialist ifumaneka okanye i-orchestrator ifumaneka, i-guard ifumaneka. I-logs ifumaneka i-error, i-alerts, kunye ne-optional isebenza i-fallback logic - ezifana ne-retrying kunye ne-agent eyahlukileyo okanye ukuguqulwa kwe-safe default response. Kwimveliso, le nqakraza ayiyi-non-negotiable. Iingxaki ziya kuthatha. I-GPT inokufumana iingxaki. I-API ingaba i-timeout. Ufuna i-layer ebonakalayo ukuba uyazi ukuba iingxaki ziyafumaneka xa izicwangciso zihambe. Umzekelo: I-GPT ibonelela i-JSON ebhalisiweyo. I-guard ibonelela, ibhalisele i-problem, ibhalisele i-Slack, kwaye ibhalisele i-fallback. { "error": "Invalid JSON from Summarizer", "fallback_response": "Sorry, I couldn’t process your request. Forwarding it to a human operator." } Observer Agent: the System’s Black Box log, ukuhlaziywa, kwaye ukunika ukubonisa into efanelekileyo. Purpose: Le mveliso akufutshane - akufutshane kuphela. Ufuna ukucacisa ixesha elide le mveliso eyahlukileyo? Yintoni iinkcukacha i-Memory uqhagamshelane? Yintoni i-Guardi kufuneka uqhuba i-fallback? I-Observer ibhokisi yonke into efuna - i-Supabase, i-Amplitude, i-Segment, okanye into eyahlukileyo. Ndicinga njengoko i-black box kwi-airplane. Ufuna ukuba awunayo, kodwa xa kukho into yokukhangisa, ke yindlela kuphela ukufumana into efanelekileyo. Ngaphandle kwe-Observer, unxibele. Ngoko, ufumane umfanekiso epheleleyo. Egress Agent: the Final Touch Ukucacisa inkqubo ngokucacileyo. Purpose: Le mveliso ithatha i-output ekugqibeleni, i-formats, kwaye ibonelela apho kufuneka uye - kwi-user, i-API, i-webhook, i-Slack, nayiphi na. Akukwazi ukuthatha imibuzo. Oku kuphela i-response yaye ibonelela. Enye yeengxaki ezininzi ziquka ukufumana le step. Ngaphandle umxholo we-egress ekhethekileyo, ungenza ukufumana idatha ebandayo, i-logs ebandayo, okanye imibuzo kwindawo ebandayo. { "status": "success", "reply": "We’ve reviewed your case. Thank you for reaching out!" } Zonke le mveliso i-building block. Uyakwazi ukuqhuba kwi-parallel, ukuguqulwa kwakhona ngokuzimela, ukusetyenziswa kwakhona kwi-chain. What’s Next? How to Start Using AAC in Your Projects Yintoni? Indlela yokuqala ukusetyenziswa kwe-AAC kwiiprojekthi yakho Ukuba uqhagamshelane le nqaku, iimfuno iye uyakufumana iimeko ezifanayo ezifanayo. inkqubo yakho ikakhulu. Ukusetyenziswa kwimeko ezininzi. I-logic ibandakanya ngaphezulu. I-agents iye "intelligent" - kodwa inkqubo ibandakanya. Yintoni i-MVP efanelekileyo kwidlulileyo kwidlulileyo. Ingaba ufuna ukongeza isakhiwo esitsha? Yinto ingxaki. Ukukhangela ukuba i-GPT iye yenza into emangalisayo? Akukho umnxeba apho kwenziwe. Yinto ngexesha ukuqala ukufumana - kukho indlela enye? Yintoni. Ukusetyenziswa kwe-AAC kunceda ukuguqulwa kwindlela yakho. It is a shift from magic to engineering. Kwaye indawo yokuqala ayikho ikhowudi. Oku kuthetha kwinkqubo yakho ngokufanelekileyo. Ngokuqhelekileyo. Akukho izibonelelo. Qinisekisa i-logic yakho njenge-chain yeengcali ezihambelana. Umntu akufumana i-input. Umntu wabanjwa iinkcukacha. Umntu wabanjwa ukuba uya. Emva koko i-executor. Emva koko - i-memory. Emva koko, umntu akufumana iingxaki. Ekugqibeleni, umntu akhawunti into. Yinto yeengxaki yakho yokuqala, i-AAC-style. Ngoku, kunjalo, zonke iingxaki ziyi-implicit okanye zihlanganisa ngokugqithisileyo ngexesha elininzi ukuba akukho umntu uyazi ukuba yintoni. I-step yokuqala yokwenene kukwenza iimpawu eziqhelekileyo. Kwakhona nje ngokucacileyo. I-Sketch them out. Give them names. Start asking the right questions: who is responsible for orchestration here? Where is memory? How do I know if things actually worked? What happens if one block fails? Xa uyenze oku, i-value ye-modularity ibonakala. Xa yonke into ifakwe kwindawo eyodwa, ingxaki elinye ibandakanya inkqubo epheleleyo. Kodwa xa izinto zihlanganisa njenge-agents ebandayo, iingxaki zihlanganisa - kwaye inkqubo ibhale. Yeyona yoko kwi-AAC, zonke i-agents zihlanganiswa, zihlanganisa nge-contracts, kwaye ayikho ukunceda kwezinye. Kuyinto njenge-microservices kwi-pipeline elinye. Kukho kuphela elula. Kwaye ngokukhawuleza. Kwaye umdla waqala. Uyakwazi ukufumana ukuba unako ukusetyenziswa kwicandelo ye-system. I-agent ye-classifier efana ne-flow elinye ingasetyenziswa kwisiqingatha. I-memory block ingasetyenziswa kwiinkwenkwezi ezininzi. Uyakwazi ukuvelisa i-orchestrator kwi-intelligent - ufumane i-agent ye-trigger. Oku akayi-fantasy. Oku yi-baseline maturity ye-AAC. Kwi-no-code platforms efana ne-n8n, oku kwenza i-superpower epheleleyo. Ngaba uqala ukwakha i-library yeengxaki. Wonke enye i-subflow kunye ne-contract ebonakalayo. Ufuna ukwakha i-bot entsha? Nceda uqhagamshelane neengxaki ezifunekayo, ngexesha elidinga. Ufuna ukuguqulwa i-block? Yenza oku ngaphandle kwexabiso. Ufuna ukuhanjiswa kwimveliso okanye ukucacisa apho izinto zikhuba kakhulu? Uyaziyaziya - I-Observer ibhokisi yonke into, i-Guard ibopha zonke iingxaki. Ufuna ukufumana i-system njenge-engineer, akukho umntu omnqweno ukuyifaka i-code. Yintoni le nceda. Uya kubonisa. Uya kubuyisa i-chain yeengxaki ze-GPT kwi-architecture yokweneneyo - into efanelekileyo, i-scales, i-logs, i-evolution. Kwaye ukuba unayo ixesha ukuguqulwa yonke into ngexesha elandelayo - ke kakuhle. Ukuqala nge-flow elinye. Umgangatho elinye. Ixesha elifanelekileyo apho uthetha, "Ndla, nceda nqakraza indlela entsha apha. Nceda ukwahlula iingxowa. Engeza i-control." Kuyinto ixesha ukuqala ukwakha kunye ne-AAC. Kwaye iziphumo ziyafumaneka - akufunayo ukuba ushiye.
