Usuku lokugqibela, i-Ship Iintsuku ezininzi ze-engineering sprints - kwiiveki elinye, kunye ne-10 iingxaki ze-Membrane Agent ezisebenza ngexesha elifanelekileyo. Ndiyathanda indlela yokwakha i-pipeline ebonakalayo. 1,000 API integrations I-Universal Membrane I-Membrane Universe yi-library yethu ye-pre-built integration knowledge - yonke into enomdlavuza okanye umdlali kufuneka uqhagamshelane ne-API ze-external. Kukho iintlobo ezininzi zeemveliso ngaphakathi, kodwa kule nophuhliso siqedela kwiiyure ezimbini: I-Connectors, ebonakalayo indlela yokuhambisa kwi-API ye-external (i-authentication nge-OAuth2, i-API keys, njl, kunye ne-data collections kunye neengxaki) I-Action Packages, eziquka iinkonzo zeengxaki ze-API ezisetyenzisiweyo (isib. "Create a Slack message", "List GitHub repos") ukuba ama-agents kunye ne-workflows zithunyelwe. Kwi-action packages, sincoma ukuba ziyafumaneka - sinikeza iingxaki ezininzi ezivela kwi-80% yokusetyenziswa okuzenzakalelayo. Kwixesha elinye, abasebenzisi zibonise iingxaki ze-ad-hoc nge-self-integration. Ukwakhiwa yonke ukuhlanganiswa ngamanzi kufuneka i-developer — uphando i-docs, ukufumana i-auth, ukuvelisa i-client, ukuyifaka i-tests. Kule izinga, i-1,000 ukuyifaka iya kuthatha umntu omnye ngonyaka lokufanelekileyo. Siza kusetyenziswa kwi-LLMs ukuvelisa oku ukususela kwiintsuku ezincinane ze-gpt-3.5, kodwa kuxhomekeke ad-hoc. 30–60 minutes I-Membrane Agent uyazi indlela yokusebenza kunye ne-platform yethu. . Siye zibonise i-batch pipeline yokusebenzisa iisicelo ezininzi ngokuzenzakalelayo. We saw the opportunity to industrialize it Ukwakhiwa kwePipeline I-pipeline ine-phase ezimbini, ngamnye ebizwa nge-batch script yayo. I-Phase 1 ibhalisele ne-authentication - ingxenye enzima ye-integration. I-Phase 2 izihlangu zokusebenza ezifanelekileyo. Zonke izihlangu zihlanganisa isakhiwo esifanayo: ukufumana i-apps efanelekileyo, ukufumana i-agent ye-AI efanelekileyo, ukuvalwa iziphumo, ukuhlaziywa into efanelekileyo, ukufaka into efanelekileyo. Isinyathelo 1 - Ukubhalisa (build connectors) I-script ithatha i-step yokuqala: ukulungiselela i-auth ngalinye le app. How it works: Yenza zonke i-apps evela kwi-API yethu, i-filter kwi-apps ngaphandle kwe-connector Kule app (kuvelisa ukuya ku-10 ngexesha elifanayo), it: Yenza i-connector record kwi-Membrane Yenza i-agent session kwi-engine yethu I-Spawns ye-Membrane Agent ye-Membrane ye-Claude Ukubonisa i-agent ye-connector yokusetyenziswa - i-agent uyazi indlela yokusebenza kunye ne-Membrane ukusuka kwi-system prompt yaye indlela yokwakha i-connectors nge-pre-loaded skills, ngoko umxholo we-user yi-app name kunye ne-URL Ukukhangisa ukuba i-agent ifumaneke (~2.5 imizuzu kumgangatho) I-Validates i-result vs. i-schemes yethu - le i-feedback loop yinto ebalulekileyo kwi-agents, njengoko zithembisa ngokufanelekileyo xa i-validation ibonakaliswa If valid: ifumanisa i-connector kwaye ibonisa ifumaneka ifumaneka: ifumanisa i-app ukuze ifumaneke ngqo What Membrane Agent actually does inside each session: Okokuqala, umqhubi usebenzisa iimveliso ukufumana i-API yedokumentation ye-app. I-Docks ibonisa ukuba i-API isebenzisa i-OAuth2, i-API keys, i-Basic auth, okanye nayiphi na into, kwaye ibonise zonke iiparametre ze-auth ezinxulumeneyo - i-Client ID / iifolda ze-secret, i-scope, i-token URLs, iimveliso. web search web fetch Emva koko lula i-API client enikezela ngokufanelekileyo i-credentials kwi-requests, ivumela i-test function yokubhalisa i-connection, kwaye ngokufanelekileyo inikeza i-HTTP requests kwi-API yokubhalisa ukuba iyatholakala kunye nokuphendula ngokufanelekileyo. Okugqibela, isebenzisa izixhobo zeMembrane ukubhalisa yonke isakhiwo kwiphepha. Le nqubo epheleleyo , kwaye i-agent ithatha ngokugqibeleleyo. 2.5 minutes per app Yenza i math: Kwaye malunga neengxubevange ze-10 ziye ziye ziye zibonwa kwaye ziye zibonwa ngexesha elinye imizuzu eziliqela - ngaphandle kokucoca umntu omnye. Kwaye i-10 yintoni ndingathanda ngexesha elide - i-concurrency iyakwazi ukuvelisa kwaye kunokuba kwenziwe ngaphezulu. 10 agents, ~2.5 minutes each, running in parallel Wonke i-agent ibhalisele i-connector (okanye i-action package) ngexesha lokusebenza. Sinikezela ngexesha elinye ngexabiso ngexesha lokusebenza ukuze akufuneka ukuchithwa kwe-context window - i-session entsha ngexabiso lokusebenza ngexabiso ibhalisele ngexabiso. Phase 2 - Imisebenzi (ukwakha iipakheji) Emva kokufaka i-app, i-auth ifanelekileyo kwi-phase yesibini: ukuvelisa iingxaki ezininzi ezibonisa ukuhlanganiswa. Le script ithatha zonke i-app eyenza i-connector kwaye yenza i-action pack for it. I-pattern ibonisa i-Phase 1. I-script i-filter kwi-apps eyenza i-connector kunye ne-auth kodwa ayikho i-package, ke i-agent yenza i-agents ngamnye. Yonke i-agent ibonelela i-ID yayo ye-connector kwaye ibonelela ukuba isebenze i-package. I-script ibonelela kwi-API ye-app, ibonise i-endpoints ezininzi ezidumileyo kwaye ibonise i-definitions ye-action - ephelele kunye ne-input schemes, i-API request configuration, i-output schemes, kunye ne-optional guidelines for non-obvious behavior. Emva kwe-validation (ukubonisa i-package ngokwenene i-actions), ibon Ukucinga Ndiyathanda njani inkqubo epheleleyo xa ufake i-zoom: Imibuzo Technical Key Ukusuka kwi-5-10, sisebenza Nazi yintoni iye yenza le umsebenzi ngokufanelekileyo: ~100 apps per batch run Ukucaciswa kweSession Zonke iingxowa ze-agent zihlanganisa kwi-cloud yethu, nangona i-agents zithunyelwe kwi-locally ngexesha le-batch builds. I-script ikhiqiza iingxowa kwi-platform yethu kwaye emva kwe-agents eyahlukileyo, kwaye i-synchronizes zonke iingxowa ze-conversation. Yintoni, sinokukwazi ukubuyekeza zonke iimpazamo ze-AI ngokusebenzisa i-UI ye-console yethu, ngexabiso njengoko i-cloud-hosted session. Kwakhona sinokufumaneka okanye ukutshintsha i-session ye-cloud xa kufuneka. Ukusabela & Ukusabela Imibuzo Kukho zonke i-app angakwazi ukuzenzakalelayo. I-script ithatha iingxaki ngokugqithisileyo: Ukuhlaziywa kwe-schema: Emva kokugqibela kwe-agent, sincoma imiphumo kwi-SDK yethu. Ukuba akuyona (ukungabikho kwifolda ezifunekayo, isakhiwo esikhoyo), i-app iyafumaneka. I-APIs ye-Dead: I-agent ibizwa ukuba ushiye i-auth embalwa kwaye ibonise ngoko ukuba i-API ayikho. Ezi zithunyelwe. I-Timeouts: Ukuba i-Cloude ibonakala kwi-API emangalisayo kakhulu (ngaphandle kokuba akuyona ngokufanelekileyo), i-session ifumaneka njenge-Failed kwaye inokufanelekileyo. Yintoni iindawo ezifanayo: iingxaki ziye zithunyelwe kwimveliso. Xa i-agent ibonakaliswa kwi-app, sinqathwe i-session ukuze ufumane ukuba yintoni iingxaki kwiingcebiso ye-agent? I-API emangalisayo? I-Documentation engabonakali? Sithunyelwe inkcebiso esisiseko, uqhutywa kwakhona, kwaye zonke iintlobo ziye kubaluleke ngexesha elidlulileyo. Izifundo ye-Agent Yinto ebalulekileyo: i-agent ayisuka ukusuka kwakhona ngalinye ngalinye i-API. I-system prompt yayo ifakwe kwizinto ezininzi zenzululwazi: I-Membrane Platform Overview (Yintoni i-Membrane, njani i-framework) I-Connector building skill (i-proprietary step-by-step workflow yokusebenza kwe-auth. ukucacisa i-auth type, ukucacisa i-auth-type-specific docs, ukucacisa iiparametre, ukucaciswa kwe-API client, ukucaciswa kwe-test) uqeqesho we-OpenAPI (ukwazi ukufumana kunye nokufunda ngokubanzi i-OpenAPI specs ngaphandle kokufaka i-schemes epheleleyo kwi-context), Iingxelo zokusebenza eziqhelekileyo ze-core connector functions. I-Agent Framework yethu inikezela ukuxhaswa kwezemvuno kwi-on-demand ngexesha le-session, kodwa kwi-batch processing sinokufumaneka ukuba ukuxhaswa kwezemvuno eziphambili ngqo kwi-system prompt isebenza kakuhle. I-LLMs ayikho kwimvuno ze-self-load ngokufanelekileyo kwi-100% kwimeko, kwaye kulesi isilinganiso ufuna ukuxhaswa kwe-flexibility. Yintoni i-agent inokufuneka ngokubanzi iimveliso ze-platform yethu ngaphambi kokufunda i-API ye-target. I-user prompt i-minimal - nje i-app name kunye ne-URL. I-Manual Layer Zonke into ayikho ngokupheleleyo okuzenzakalelayo - kwaye oku yi-design. Zonke into akufuneka umntu: Iingxaki ze-Edge: Ezinye i-API ze-undocumented kodwa ezisebenzayo. Thina baye ziye zibonakalisa ngexesha lokuhlola kwaye zithunyelwe ngqo. Ukubuyekezwa Kwekhwalithi: Sinikezela kwi-agent sesiyoni kwi-console yethu, ikakhulukazi kwi-apps apho ukuguqulwa kwangapheli. Ayikho iziqinisekiso ezininzi: Kwixesha elide, i-agent ayifumaneka nge-API ezininzi. I-API ibonise ukuba ziyafumaneka kwaye ukuba i-auth ifumaneka ngokufanelekileyo, kodwa ayifumaneka kwi-OAuth efanelekileyo. Thina ukwakha ngokugqithisileyo i-browser automation ye-auto-test account generation ukuze kufake le ingxaki. Yintoni i I-Membrane Universe ithathwe ngokubanzi kwiiveki ezidlulileyo, ukusuka kwi-niche kunye ne-apps ezizodwa - i-API ze-old-school, i-systems ezincinane ezidokumented. Ukwakhiwa kwe-browser automation ye-automatic signup kunye ne-OAuth flows ukuze ama-agents angafumana i-end-to-end integrations. Ixesha elide: ukugcinwa okuqhubekayo. I-API ifumaneka, i-endpoints zithunyelwe. I-agents efanayo ziye ziye ziye ziye ziye ziye ziye zithunyelwe. Imibala elikhulu yinto: I-AI agents ayikho kuphela ama-coding assistants eziza kukunceda ukubhala iifundo ngokukhawuleza. . Ukucinga kubo kwinkinga elifanelekileyo, ukunika izixhobo ezifanelekileyo kunye nezifundo, kwaye ziyafumaneka ukuvelisa izinto kwi-scale eyayifunayo nje. Sikhokelela kwi-1000 API zethu, kwaye ziye zithunyelwe. infrastructure builders
