TL;DR. Agentic AI ya wuce daga gwaji zuwa samarwa a bankunan duniya. Kashi 70% na cibiyoyi suna amfani da shi; ɗaya cikin biyar ne kawai ke da ƙwararre na mulki. Abokan gaba masu cin gashin kansu suna aiki da saurin injin, gado na COBOL an rubuta shi don zato na sarrafawa-tari na 1960s, kuma EU AI Act ta Agusta 2026 tana nesa da makonni goma sha biyu.

Mahimman Bayanai

Shekarar Da Agentic Engineering Ya Zama Makawa. Tattaunawa game da AI a cikin sabis na kuɗi, har zuwa kwanan nan, abubuwa biyu maƙwabta amma mabanbanta ne suka mamaye shi: hanyoyin sadarwa na tattaunawa ta generative (masu taimako amma suna da iyaka), da kuma…
Halin Karɓa A Harkar Banki. Cikakken hoton a bayyane yake.
Hanyoyi Guda Uku Na Haɗari Da Dole Ne Bankuna Su Fahimta. Kafin kowace tattaunawa ta tsari, hankalin hukumar gudanarwa ya kamata ya karkata akan haɗari guda uku waɗanda ke da alaƙa da tsarin agentic kuma za su zo da wuri fiye da yadda yawancin bankuna suka tsara.
Shekarar Da Agentic Engineering Ya Zama Makawa. Tattaunawa game da AI a cikin sabis na kuɗi, har zuwa kwanan nan, abubuwa biyu maƙwabta amma mabanbanta ne suka mamaye shi: hanyoyin sadarwa na tattaunawa ta generative (masu taimako amma suna da iyaka), da kuma…

Agentic Engineering don Bankuna: Tsarin 2026 don C-Suite da Injiniyoyin da Za Su Gina Shi

Agentic AI ya wuce daga gwaji zuwa samarwa a bankunan duniya. Kashi saba'in cikin ɗari na cibiyoyi suna amfani da shi har zuwa wani mataki; ɗaya cikin biyar ne kawai ke da ƙwararre na mulki. A halin yanzu, abokan gaba masu cin gashin kansu suna aiki da saurin injin, gado na COBOL wanda sabbin tsarin dole ne su yi aiki da shi an rubuta shi don zato na sarrafawa-tari (batch-processing) na shekarun 1960, kuma lokacin da ya ƙare na EU AI Act na haɗari mai tsanani yana nesa da makonni goma sha biyu. Wannan shi ne matsayi na injiniya da shugabanci da banki ke buƙatar kiyayewa.

Muhimman Abubuwan Kula

Canji daga vibe coding zuwa spec-driven development ba buri ba ne kawai yanzu. Andrej Karpathy, wanda ya ƙirƙiri kalmar "vibe coding" a watan Fabrairu na 2025, ya amince shekara guda bayan haka ⧉ cewa lokacin yana ƙarewa kuma sabon tsoho ga ƙwararru shi ne agentic engineering — tsara wakilai (agents) akan cikakkun bayanai (specifications) tare da kulawar ɗan adam.

Karɓar banki na gaske ne kuma yana sauri. Kashi 70% na kamfanonin banki ⧉ sun ba da rahoton amfani da agentic AI har zuwa wani mataki (Kashi 16% a cikin samarwa, kashi 52% a cikin gwaji, EY 2026); Kashi 44% na ƙungiyoyin kuɗi za su yi amfani da shi a wannan shekarar — ƙaruwa ta sama da 600%+ shekara-shekara a cewar Wolters Kluwer.

Gudanarwa bai bi sauri ba. Binciken Deloitte na State of AI 2026 ya gano cewa kamfani ɗaya kawai cikin biyar ne ke da ƙwararren tsarin gudanarwa don wakilan AI masu cin gashin kansu. Binciken Deloitte na MIT AI Risk Database ya gano fiye da haɗari 350 ⧉ waɗanda za su iya tasowa daga halayyar cin gashin kai ko na agentic.

Yanayin barazana ya zama na masana'antu. Anthropic ta bayyana a watan Nuwamba na 2025 cewa rukunin da ke samun goyon bayan gwamnatin Sin GTG-1002 ya yi garkuwa da Claude Code don gudanar da aikin leƙen asiri na kansa akan kusan maƙasudai 30, inda AI ke gudanar da kashi 80-90% na ayyukan dabaru da kanta. Flashpoint ta lura da hawan kashi 1,500% a cikin tattaunawa ta haramtacciyar hanya mai alaƙa da AI ⧉ tsakanin Nuwamba da Disamba na 2025 kaɗai.

Tsohon tsarin gado shi ne babban cikas na shiru. Kudin IT na ayyukan kuɗi yana cinye kashi 70-75% ta hanyar kula da tsoffin tsarin gado, kashi 63% na bankuna har yanzu suna dogaro da lambobin da aka rubuta kafin shekarar 2000, kuma yawancin bankuna sun ba da rahoton mutum ɗaya ko biyu kawai a cikin gida waɗanda za su iya kula da COBOL da manyan dandamalinsu ke gudana akai. Agentic AI yanzu shi ne babban tsarin da ake amfani da shi don cike wannan gibi.

Tarihin ƙa'idodi yana haɗuwa. A ƙarƙashin EU AI Act, 2 ga Agusta 2026 yana kunna cikakken aiwatarwa don tsarin AI masu haɗari sosai (Annex III a sarari ya haɗa da credit scoring da creditworthiness assessment). DORA riga yana aiki. An tsawaita SR 11-7 a cikin aikin mai sarrafawa don rufe LLMs da tsarin agentic. Tara don keta doka ya kai Yuro miliyan 35 ko kashi 7% na kuɗin shiga na shekara-shekara na duniya.

Kulawar ɗan adam ba ra'ayi ɗaya ba ne. Bambanci tsakanin HITL (Human-in-the-Loop, inda wakilin ba zai iya aiwatarwa ba tare da amincewar ɗan adam ta musamman ba) da kuma HOTL (Human-on-the-Loop, inda wakilin ke aiwatarwa da kansa ƙarƙashin kulawar ɗan adam) yanzu shi ne tsarin aiki don bin umarnin sashe na 14 na EU AI Act, kuma kowane wakili mai babban haɗari yana buƙatar bayyana takamaiman matsayi akan wane samfurin ne ke aiki.

Yawancin wakilai za a saya ne, ba gina su ba. Gudanar da haɗarin ɓangare na uku (TPRM) ƙarƙashin DORA shi ne babban ƙalubale da ba a cika lura da shi ba a cikin 2026. Masu siyarwa (vendors) za su ba da mafi yawan ƙarfin agentic da bankuna ke tura wa; alƙawarin ƙa'ida ya kasance a kan banki, kuma mafi yawan kwangilolin masu siyarwa na yanzu ba za su iya cika buƙatun takaddun Sashe na 13 ba.

Agentic engineering ba "ChatGPT tare da sabobin MCP" ba ne. Matsayi ne na mallakar tsarin gaba-ɗaya na tafiyar cibiyar — tafiye-tafiyen abokin ciniki, yanayin rayuwar ma'amala, control plane, audit substrate, quantum-safe cryptographic foundation — wanda aka gina kuma yake gudana ta sashin injiniya na cibiyar kanta, ba wanda aka damƙa wa chatbot ba.

Shekarar Da Agentic Engineering Ya Zama Makawa

Tattaunawa game da AI a cikin sabis na kuɗi, har zuwa kwanan nan, abubuwa biyu maƙwabta amma mabanbanta ne suka mamaye shi: hanyoyin sadarwa na tattaunawa ta generative (masu taimako amma suna da iyaka), da kuma tsarin Retrieval-Augmented Generation da aka ɗora akan bayanan kamfanoni (masu amfani, kuma suna da iyaka). Abin da ya canza tsakanin ƙarshen 2025 da farkon 2026 shi ne cewa rukuni na uku — wakilai masu zaman kansu (autonomous agents) waɗanda ke tsara labarai, aiwatarwa, da kuma kammala ayyuka masu matakai da yawa tare da iyakantaccen kulawar ɗan adam — ya tashi daga nunin fasaha zuwa ainihin aiki, kuma ya shiga lokaci guda cikin kamfanoni da kuma masu aikata laifuka.

Andrej Karpathy, wanda ya ƙirƙiri kalmar "vibe coding" a watan Fabrairu na 2025 ⧉, ya kalli yadda kwararrun injiniyoyi suka wuce wannan mataki a cikin shekara ta gaba. Sabon kalmar da ya yi amfani da ita — "agentic engineering" — yanzu ita ce kalmar aiki a duk faɗin masana'antar. Ainihin wannan canjin mai sauƙi ne: a cikin manyan ayyukan software a 2026, injiniyoyi ba su da rubuta lambar da kansu kashi 99% na lokaci. Suna tsara wakilan da ke yin hakan, yayin da su kuma suke aiki a matsayin masu sa ido. Aikin ba rubuta haruffa a cikin edita ba ne kuma; samar da bayanai (specifications) ne waɗanda ke taƙaita abin da wakilan za su iya samarwa, tsara hanyoyin tabbatarwa (verification gates) waɗanda dole ne sakamakon ya wuce, da kuma kula da yanke shawara na tsarin gine-gine (architectural decisions) waɗanda wakilan ke aiwatarwa.

Wannan canjin yana sauti kamar tattaunawar ƙungiyar injiniya ce. A harkar banki ba haka ba ne. Tattaunawa ce a matakin hukumar gudanarwa (board-level), saboda irin wannan ikon na agentic da ke sake fasalin yadda ake samar da lambar cikin gida shi ma yana sake fasalin yadda abokan gaba na waje ke aiki, yadda masu sarrafawa ke tsammanin za a gudanar da sa ido, da kuma yadda ake ayyana iyakokin cibiyar. Bankin da ba shi da takamaiman matsayi akan agentic engineering a ƙarshen shekarar 2026 ba bankin da ya guje wa tambayar ba ne. Banki ne wanda masu ba shi kayayyaki (vendors), abokan gaba, da masu sarrafawa suka riga suka amsa masa tambayar.

Halin Karɓa A Harkar Banki

Cikakken hoton a bayyane yake. Dangane da binciken da aka tattara daga bincike da yawa na 2026, kashi 70% na shuwagabannin banki ⧉ sun ba da rahoton cewa kamfanoninsu sun riga sun yi amfani da agentic AI har zuwa wani mataki. Gartner yana hasashen ⧉ cewa a ƙarshen shekarar 2026 kusan kashi 40% na duk kamfanonin sabis na kuɗi za su yi amfani da wakilan AI ta wata hona. Kashe kuɗi akan AI a cikin sabis na kuɗi yana kan hanyar kaiwa dala biliyan 67 nan da 2028 (IDC). McKinsey ta kiyasta cewa agentic AI na iya dawo da sa'o'i 10-12 a kowane mako ga manajojin dangantaka (relationship managers) a harkar banki.

Hoton aiwatarwa ba shi da ban ƙarfafa sosai. KPMG ta ba da rahoton ⧉ cewa kashi 99% na kamfanoni suna shirin tura wakilai masu zaman kansu (autonomous agents) cikin samarwa amma kashi 11% ne kawai suka yi hakan. EY ta gano cewa kashi 34% na shuwagabanni sun fara amfani da wakilan AI kuma kashi 14% ne kawai suka aiwatar da su gaba ɗaya. Forrester ta gano cewa kashi 57% na ƙungiyoyi sun yi imanin cewa ba su da ƙarfin cikin gida don amfana da agentic AI. Gibi tsakanin niyya da aiwatarwa ba batun tallace-tallace ba ne. Ainihin nuni ne na injiniya, gudanarwa, da kuma aikin al'adu wanda har yanzu ba a yi ba.

Hukumar Kula da Ayyukan Kuɗi ta Burtaniya (FCA) ta bayyana damuwa a fili ⧉ game da saurin tura wakilan AI wanda ya wuce ƙwararren gudanarwa — wani tashin hankali da Babban Jami'in Bayanai na FCA Jessica Rasu ya bayyana a matsayin haɗari na kusa ga masu amfani da kayayyaki (retail consumers). McKinsey ita ma ta daban ta yi gargaɗin cewa bankunan da suka gaza daidaita tsarin kasuwancinsu ⧉ suna cikin haɗarin rasa har zuwa dala biliyan 170 na ribar duniya nan da shekarar 2030. Duk abubuwan lura guda biyu daidai suke a lokaci guda. Tambayar ba ita ce ko za a motsa ba; ita ce yadda za a motsa tare da amincin aiki da gudanarwa wanda ƙa'idodin sabis na kuɗi koyaushe suke buƙata, kuma tsarin agentic ya sa su zama mafi bayyane.

Hanyoyi Guda Uku Na Haɗari Da Dole Ne Bankuna Su Fahimta

Kafin kowace tattaunawa ta tsari, hankalin hukumar gudanarwa ya kamata ya karkata akan haɗari guda uku waɗanda ke da alaƙa da tsarin agentic kuma za su zo da wuri fiye da yadda yawancin bankuna suka tsara.

1. Abokin Gaba Mai Zaman Kansa (The Autonomous Adversary)

Babban ci gaba mafi rikitarwa a cikin 2026 shi ne amfani da agentic AI a ɓangaren hari. A watan Agusta na 2025, Anthropic ta bayyana wani nau'in aiki da ta kira vibe hacking ⧉: masu laifin intanet suna amfani da agentic AI don gudanar da hadadden hari a babban mataki, inda AI ke cikin aikin bincike (reconnaissance), satar bayanan sirri (credential harvesting), shiga hanyar sadarwa (network penetration), da kuma nazarin bayanan da aka sata. A watan Nuwamba na 2025 ⧉, Anthropic ta bayyana cewa ta dakatar da wani kamfen na wani rukuni mai samun goyon bayan gwamnatin Sin (wanda aka kira GTG-1002) wanda ya yi garkuwa da Claude Code don gudanar da aikin leƙen asiri na kansa akan kusan maƙasudai talatin na tsaro, makamashi, da fasaha, inda AI ke tafiyar da kashi 80-90% na ayyukan dabaru kuma yana aiki a kan dubu-dubu na buƙatu a kowane daƙiƙa — saurin da ba zai yiwu ba ga mutum.

A watan Janairu na 2026, an keta tsaron Step Finance — mai kula da DeFi a kan Solana — ta hanyar da ta mayar da kutse a cikin na'ura zuwa asasar dala miliyan 27-30 saboda wakilan kasuwanci na AI na kamfanin suna da izinin aiwatar da manyan canje-canje na kuɗi ba tare da amincewar ɗan adam ba. Maharin ya yi amfani da dabarun injiniyan zamantakewa (social engineering) akan AI kanta, yana da'awar cewa yana gudanar da shirin gano kura-kurai mai izini (bug bounty). Darasi ⧉ a nan ba wai cewa AI ba shi da aminci gaba ɗaya ba ne; a'a, shi ne cewa wakilin AI wanda ke karɓar iƙirarin izini ba tare da tantancewa ba rauni ne ga tsaro.

Cikakken yanayin shi ne abin da dole ne bankuna su fahimta. Rahoton Flashpoint na 2026 na Global Threat Intelligence ya gano hawan kashi 1,500% na tattaunawa ta haramtacciyar hanya mai alaƙa da AI ⧉ tsakanin Nuwamba da Disamba na 2025, tare da maharan da ke haɓaka tsarin cin gashin kai waɗanda ke tattara bayanai (scrape data), canza abubuwan more rayuwa (rotate infrastructure), daidaita saƙonni, da kuma koyo daga yunƙurin da ya gaza ba tare da kulawar ɗan adam na ci gaba ba. Jamie Dimon na JPMorgan ya bayyana a fili ⧉ cewa fa'idar farko ta wannan fasaha tana tafiya ne ga masu kai hari (offence), ba masu tsaro ba (defence). Ma'anar wannan tana da wuyar sha'ani: bankin da ke gudanar da ayyukan tsaro na gargajiya akan abokan gaba na agentic, a tsari, yana cikin matsayin ɗan wasan dara (chess) wanda aka ba abokin hamayyarsa kwamfuta.

2. The Code-Quality Regression

Hanya ta hui tana faruwa ne a cikin gida kuma tana da shiru. Lambar da LLM ta samar, idan babu tsarin bayanai na musamman da kuma tantancewa mai tsanani, tana fitowa da lahani a kan mafi girma fiye da lambar da mutum ya rubuta. Binciken SonarQube na LLMs biyar na gaba ⧉ da ke samar da lambar Java ya gano cewa fiye da kashi 70% na raunin da aka gano a cikin sakamakon Llama 3.2 90B an sanya su a matsayin babban haɗari na BLOCKER, tare da kusan kashi biyu cikin uku na raunin GPT-4o da OpenCoder-8B da aka sanya su a matsayin BLOCKER ko CRITICAL. Pearce et al. (IEEE S&P) sun g BEGIN_TRANSLATION

author: "contact@sebastienrousseau.com (Sebastien Rousseau)" banner_alt: "Zane na gine-ginen injiniyan agentic ga bankunan — wakilan AI na spec-driven, farantin kula da mulki, substrate mai amincin quantum, da kwararan modernisation na mainframe na gado" banner_height: "571" banner_width: "1425" banner: "https://cloudcdn.pro/stocks/images/hector-j-rivas-1FxMET2U5dU-unsplash.webp" cdn: "https://cloudcdn.pro" charset: "UTF-8" cname: "sebastienrousseau.com" copyright: "© Copyright 2007 - 2026 - Sebastien Rousseau. An kiyaye duk hakkokin." date: "17 May 2026" description: "Agentic AI ya wuce daga gwaji zuwa samarwa a bankunan duniya. Kashi 70% na cibiyoyi suna amfani da shi; ɗaya cikin biyar ne kawai ke da ƙwararre na mulki. Abokan gaba masu cin gashin kansu suna aiki da saurin injin, gado na COBOL an rubuta shi don zato na sarrafawa-tari na 1960s, kuma EU AI Act ta Agusta 2026 tana nesa da makonni goma sha biyu." format-detection: "telephone=no" hreflang: "ha" icon: "https://cloudcdn.pro/clients/sebastienrousseau/v1/logos/sebastienrousseau.svg" id: "https://sebastienrousseau.com/ha/2026-05-17-agentic-engineering-banks-blueprint-2026" image_alt: "Hoton Sebastien Rousseau a Baki da Fari" image_height: "162" image_width: "162" image: "https://cloudcdn.pro/stocks/images/sebastienrousseau.webp" keywords: "injiniyan agentic, spec-driven development, bankunan, sabis na kuɗi, AI governance, EU AI Act, DORA, SR 11-7, NIST AI RMF, ISO 42001, COBOL modernisation, mainframe, vibe coding, vibe hacking, autonomous agents, Claude Code, GitHub Spec Kit, AI Bill of Materials, AIBOM, human in the loop, HITL, human on the loop, HOTL, sarrafa haɗarin ɓangare na uku, TPRM, vendor risk, post-quantum cryptography, agent control plane, FCA, SM&CR, shirin aiki na makonni 12, Sebastien Rousseau" language: "ha-NG" layout: "report" locale: "ha_NG" logo_alt: "Tambarin Sebastien Rousseau" logo_height: "44" logo_width: "44" logo: "https://cloudcdn.pro/clients/sebastienrousseau/v1/logos/sebastienrousseau.svg" menu: "" measurementID: "G-169G4ET5HQ" name: "Sebastien Rousseau" permalink: "https://sebastienrousseau.com/ha/2026-05-17-agentic-engineering-banks-blueprint-2026" rating: "general" referrer: "no-referrer" robots: "index, follow" schema: "FAQPage, Article" seo_title: "Injiniyan Agentic don Bankuna: Tsarin Gudanar da AI na 2026" short_name: "sebastienrousseau" subtitle: "Agentic AI ya wuce daga gwaji zuwa samarwa. 70% na bankunan suna amfani da shi; ɗaya cikin biyar ne kawai ke da ƙwararre na mulki. Abokan gaba suna aiki da saurin injin, gado ya rubuta don zato na sarrafawa-tari na 1960s, kuma lokacin da ya ƙare na EU AI Act na haɗari mai tsanani yana nesa da makonni goma sha biyu." tags: "agentic AI, spec-driven development, banks, mulkin AI, EU AI Act, DORA, COBOL modernisation, HITL, HOTL, TPRM, AIBOM, Claude Code, post-quantum cryptography, ISO 20022, quantum computing, AI, Rust, open source, platform engineering" theme-color: "0, 67, 165" title: "Agentic Engineering na Bankuna: Tsarin 2026 na C-Suite da Injiniyoyin da Zasu Gina Shi" url: "https://sebastienrousseau.com/ha/2026-05-17-agentic-engineering-banks-blueprint-2026" viewport: "width=device-width, initial-scale=1, shrink-to-fit=no" last_reviewed: "2026-05-17"

RSS - The RSS feed front matter (YAML).

atom_link: "https://sebastienrousseau.com/ha/2026-05-17-agentic-engineering-banks-blueprint-2026/rss.xml" category: "Finance" docs: https://validator.w3.org/feed/docs/rss2.html generator: "Static Site Generator (SSG) (version 0.0.26)" item_description: "Cikakken tsarin 2026 don injiniyan agentic a bankuna da cibiyoyin kuɗi — wanda ya shafi canji daga vibe coding zuwa spec-driven development, HITL vs HOTL, sarrafa haɗarin ɓangare na uku (TPRM) ƙarƙashin DORA, tsarin tsari, sabunta gado (legacy modernisation), da kuma substrate mai amincin quantum." item_guid: "https://sebastienrousseau.com/ha/2026-05-17-agentic-engineering-banks-blueprint-2026/rss.xml" item_link: "https://sebastienrousseau.com/ha/2026-05-17-agentic-engineering-banks-blueprint-2026/rss.xml" item_pub_date: "Sun, 17 May 2026 06:06:06 +0000" item_title: "Agentic Engineering na Bankuna: Tsarin 2026 na C-Suite da Injiniyoyin da Zasu Gina Shi" last_build_date: "Sun, 17 May 2026 06:06:06 +0000" managing_editor: "contact@sebastienrousseau.com (Sebastien Rousseau)" pub_date: "Sun, 17 May 2026 06:06:06 +0000" ttl: "60" type: "article" webmaster: "contact@sebastienrousseau.com"

Apple - The Apple front matter (YAML).

apple_mobile_web_app_orientations: "portrait" apple_touch_icon_sizes: "192x192" apple-mobile-web-app-capable: "yes" apple-mobile-web-app-status-bar-inset: "black" apple-mobile-web-app-status-bar-style: "black-translucent" apple-mobile-web-app-title: "Agentic Engineering ga Bankuna" apple-touch-fullscreen: "yes"

MS Application - The MS Application front matter (YAML).

msapplication-navbutton-color: "0, 83, 191"

Twitter Card - The Twitter Card front matter (YAML).

twitter_card: "summary_large_image" twitter_creator: "@wwdseb" twitter_description: "Cikakken tsarin 2026 don injiniyan agentic a bankuna da cibiyoyin kuɗi — spec-driven development, HITL vs HOTL, TPRM ƙarƙashin DORA, tsarin tsari, sabunta gado, da kuma substrate mai amincin quantum." twitter_image: "https://cloudcdn.pro/clients/sebastienrousseau/v1/logos/sebastienrousseau.svg" twitter_image_alt: "Tambarin Sebastien Rousseau" twitter_site: "@wwdseb" twitter_title: "Injiniyan Agentic don Bankuna: Tsarin Gudanar da AI na 2026" twitter_url: "https://sebastienrousseau.com/ha/2026-05-17-agentic-engineering-banks-blueprint-2026"

Humans.txt - The Humans.txt front matter (YAML).

author_website: "https://sebastienrousseau.com" author_twitter: "@wwdseb" author_location: "London, UK" thanks: "Godiya ga karantawa!" site_last_updated: "2026-05-17" site_standards: "HTML5, CSS3, RSS, Atom, JSON, XML, YAML, Markdown, TOML" site_components: "Kaishi, Kaishi Builder, Kaishi CLI, Kaishi Templates, Kaishi Themes" site_software: "Static Site Generator, Rust"

excerpt: "Agentic AI ya wuce daga gwaji zuwa samarwa a bankunan duniya. Kashi saba'in cikin ɗari na cibiyoyi suna amfani da shi har zuwa wani mataki; ɗaya cikin biyar ne kawai ke da ƙwararre na mulki. A halin yanzu,…"

Agentic Engineering na Bankuna: Tsarin 2026 na C-Suite da Injiniyoyin da Zasu Gina Shi

Muhimman Abubuwan Kula

Canji daga vibe coding zuwa spec-driven development ba buri ba ne kawai yanzu. Andrej Karpathy, wanda ya ƙirƙiri kalmar "vibe coding" a watan Fabrairu na 2025, ya amince shekara guda bayan haka ⧉ cewa lokacin yana ƙarewa kuma sabon tsoho ga ƙwararru shi ne agentic engineering — tsara wakilai (agents) akan cikakkun bayanai (specifications) tare da kulawar ɗan adam.

Karɓar banki na gaske ne kuma yana sauri. Kashi 70% na kamfanonin banki ⧉ sun ba da rahoton amfani da agentic AI har zuwa wani mataki (Kashi 16% a cikin samarwa, kashi 52% a cikin gwaji, EY 2026); Kashi 44% na ƙungiyoyin kuɗi za su yi amfani da shi a wannan shekarar — ƙaruwa ta sama da 600%+ shekara-shekara a cewar Wolters Kluwer.

Gudanarwa bai bi sauri ba. Binciken Deloitte na State of AI 2026 ya gano cewa kamfani ɗaya kawai cikin biyar ne ke da ƙwararren tsarin gudanarwa don wakilan AI masu cin gashin kansu. Binciken Deloitte na MIT AI Risk Database ya gano fiye da haɗari 350 ⧉ waɗanda za su iya tasowa daga halayyar cin gashin kai ko na agentic.

Yanayin barazana ya zama na masana'antu. Anthropic ta bayyana a watan Nuwamba na 2025 cewa rukunin da ke samun goyon bayan gwamnatin Sin GTG-1002 ya yi garkuwa da Claude Code don gudanar da aikin leƙen asiri na kansa akan kusan maƙasudai 30, inda AI ke gudanar da kashi 80-90% na ayyukan dabaru da kanta. Flashpoint ta lura da hawan kashi 1,500% a cikin tattaunawa ta haramtacciyar hanya mai alaƙa da AI ⧉ tsakanin Nuwamba da Disamba na 2025 kaɗai.

Tsohon tsarin gado shi ne babban cikas na shiru. Kudin IT na ayyukan kuɗi yana cinye kashi 70-75% ta hanyar kula da tsoffin tsarin gado, kashi 63% na bankuna har yanzu suna dogaro da lambobin da aka rubuta kafin shekarar 2000, kuma yawancin bankuna sun ba da rahoton mutum ɗaya ko biyu kawai a cikin gida waɗanda za su iya kula da COBOL da manyan dandamalinsu ke gudana akai. Agentic AI yanzu shi ne babban tsarin da ake amfani da shi don cike wannan gibi.

Tarihin ƙa'idodi yana haɗuwa. A ƙarƙashin EU AI Act, 2 ga Agusta 2026 yana kunna cikakken aiwatarwa don tsarin AI masu haɗari sosai (Annex III a sarari ya haɗa da credit scoring da creditworthiness assessment). DORA riga yana aiki. An tsawaita SR 11-7 a cikin aikin mai sarrafawa don rufe LLMs da tsarin agentic. Tara don keta doka ya kai Yuro miliyan 35 ko kashi 7% na kuɗin shiga na shekara-shekara na duniya.

Kulawar ɗan adam ba ra'ayi ɗaya ba ne. Bambanci tsakanin HITL (Human-in-the-Loop, inda wakilin ba zai iya aiwatarwa ba tare da amincewar ɗan adam ta musamman ba) da kuma HOTL (Human-on-the-Loop, inda wakilin ke aiwatarwa da kansa ƙarƙashin kulawar ɗan adam) yanzu shi ne tsarin aiki don bin umarnin sashe na 14 na EU AI Act, kuma kowane wakili mai babban haɗari yana buƙatar bayyana takamaiman matsayi akan wane samfurin ne ke aiki.

Yawancin wakilai za a saya ne, ba gina su ba. Gudanar da haɗarin ɓangare na uku (TPRM) ƙarƙashin DORA shi ne babban ƙalubale da ba a cika lura da shi ba a cikin 2026. Masu siyarwa (vendors) za su ba da mafi yawan ƙarfin agentic da bankuna ke tura wa; alƙawarin ƙa'ida ya kasance a kan banki, kuma mafi yawan kwangilolin masu siyarwa na yanzu ba za su iya cika buƙatun takaddun Sashe na 13 ba.

Agentic engineering ba "ChatGPT tare da sabobin MCP" ba ne. Matsayi ne na mallakar tsarin gaba-ɗaya na tafiyar cibiyar — tafiye-tafiyen abokin ciniki, yanayin rayuwar ma'amala, control plane, audit substrate, quantum-safe cryptographic foundation — wanda aka gina kuma yake gudana ta sashin injiniya na cibiyar kanta, ba wanda aka damƙa wa chatbot ba.

Shekarar Da Agentic Engineering Ya Zama Makawa

Andrej Karpathy, wanda ya ƙirƙiri kalmar "vibe coding" a watan Fabrairu na 2025 ⧉, ya kalli yadda kwararrun injiniyoyi suka wuce wannan mataki a cikin shekara ta gaba. His revision — "agentic engineering" — yanzu shi ne kalmar aiki a duk faɗin masana'antar. Ainihin wannan canjin mai sauƙi ne: a cikin manyan ayyukan software a 2026, injiniyoyi ba su da rubuta lambar da kansu kashi 99% na lokaci. Suna tsara wakilan da ke yin hakan, yayin da su kuma suke aiki a matsayin masu sa ido. Aikin ba rubuta haruffa a cikin edita ba ne kuma; samar da bayanai (specifications) ne waɗanda ke taƙaita abin da wakilan za su iya samarwa, tsara hanyoyin tabbatarwa (verification gates) waɗanda dole ne sakamakon ya wuce, da kuma kula da yanke shawara na tsarin gine-gine (architectural decisions) waɗanda wakilan ke aiwatarwa.

Halin Karɓa A Harkar Banki

Cikakken hoton a bayyane yake. Dangane da binciken da aka tattara daga bincike da yawa na 2026, kashi 70% na shuwagabannin banki ⧉ sun ba da rahoton cewa kamfanoninsu sun riga sun yi amfani da agentic AI har zuwa wani mataki. Gartner yana hasashen ⧉ cewa a ƙarshen shekarar 2026 kusan kashi 40% na duk kamfanonin sabis na kuɗi za su yi amfani da wakilan AI ta wata hanya. Kashe kuɗi akan AI a cikin sabis na kuɗi yana kan hanyar kaiwa dala biliyan 67 nan da 2028 (IDC). McKinsey ta kiyasta cewa agentic AI na iya dawo da sa'o'i 10-12 a kowane mako ga manajojin dangantaka (relationship managers) a harkar banki.

Hanyoyi Guda Uku Na Haɗari Da Dole Ne Bankuna Su Fahimta

1. Abokin Gaba Mai Zaman Kansa (The Autonomous Adversary)

2. The Code-Quality Regression

Hanya ta biyu tana faruwa ne a cikin gida kuma tana da shiru. Lambar da LLM ta samar, idan babu tsarin bayanai na musamman da kuma tantancewa mai tsanani, tana fitowa da lahani a kan mafi girma fiye da lambar da mutum ya rubuta. Binciken SonarQube na LLMs biyar na gaba ⧉ da ke samar da lambar Java ya gano cewa fiye da kashi 70% na raunin da aka gano a cikin sakamakon Llama 3.2 90B an sanya su a matsayin babban haɗari na BLOCKER, tare da kusan kashi biyu cikin uku na raunin GPT-4o da OpenCoder-8B da aka sanya su a matsayin BLOCKER ko CRITICAL. Pearce et al. (IEEE S&P) sun gano cewa kusan kashi 40% na shirye-shiryen da LLM ta samar a cikin yanayi mai alaƙa da tsaro suna ɗauke da rauni. Yan et al. (2025) sun sanya wannan rukunin tsakanin kashi 9.8-42.1% a cikin ma'aunin gwajinsu. Wani kundin daban na Fu et al. ya gano CWEs guda 43 a cikin kayan aikin samar da lamba na AI guda uku.

Ga masana'antar da ba ta da ƙa'idodi, wannan haraji ne akan yawan aiki (productivity tax). Ga banki, haɗari ne na ƙa'ida da aiki wanda ke taruwa. Lambar da ke shiga tare da babban rauni cikin tsarin da ke sarrafa biyan kuɗi, sulhu (settlement), ko bayanan abokin ciniki ba matsala ce ta kyawun lamba kawai a sarari ba; ita ce hanyar da abokan gaba na matakin GTG-1002 za su bincika a cikin 2027 tare da irin kayan aikin agentic da suka samar da ita. Garkuwa ba ita ce ta hana lambar da LLM ta samar ba (wanda ba zai yiwu ba ta fuskar kasuwanci) amma ita ce a kewaye ta da hanyoyin tabbatarwa da tsarin bayanai na musamman don tabbatar da cewa kurakurai sun fito fili kafin a tura su. Wannan shi ne dalili na zahiri da ya sa ake ɗaukar spec-driven development cikin sauri ta ƙungiyoyin injiniya na kamfanoni waɗanda ba na fasaha ba ne na asali.

3. The Legacy Anchor

Hanya ta uku ita ce wadda bankuna suka fi fahimta, kuma canjin agentic ya sa ta zama cikin gaggawa da kuma sauƙin warwarewa lokaci guda. Fiye da kashi 70% na kamfanonin Fortune 500 har yanzu suna dogaro da mainframes, kamar yadda binciken Computer Weekly ya lura ⧉, galibi an gina su akan shekaru da yawa na COBOL da RPG da aka haɗa tare da dabarun kasuwanci na al'ada. A cikin ayyukan kuɗi musamman, tsoffin fasahohin gado suna cinye kashi 70-75% na kashe kuɗin IT na shekara-shekara ⧉. Wani binciken CIO da aka ambata a cikin nazarin masana'antar na 2026 ya gano cewa kashi 63% na bankuna har yanzu suna dogaro da lambobin da aka rubuta kafin shekarar 2000, kuma fiye da kashi 75% sun ba da rahoton cewa suna da mutum ɗaya ko biyu kawai a cikin gida waɗanda ke da ƙwarewar kula da shi.

Abin da ya canza a watan Fabrairu na 2026 shi ne zuwan kayan aikin agentic masu aminci don sabunta gado (legacy modernisation). Sanarwar da Anthropic ta yi cewa Claude Code yana iya taswirar dogaro na COBOL, rubuta takaddun ayyuka, da kuma gano haɗari ⧉ waɗanda manazarta mutane za su ɗauki watanni kafin su fitar da su — tare da irin waɗannan damar daga Microsoft (GitHub Copilot for COBOL, Watsonx Code Assistant) da kuma AWS (Mainframe Modernization tare da agentic AI) — ya rage tsadar kuɗin sabuntawa sosai. Martanin da aka gani a farashin hannun jarin IBM (faduwar kashi 13% a ranar sanarwar) ya kasance siginar kasuwa mai ma'ana ta gaske. AI yanzu yana wakiltar kusan kashi ɗaya cikin uku na jarin sabunta kamfanoni, kuma fiye da kashi 75% na kamfanoni suna amfani da AI a cikin dabarun sabuntawarsu. Katoton gado, a karon farko, ya zama matsalar injiniya da za a iya warwarewa maimakon matsala ta tsararraki.

Dalilin Da Ya Sa Vibe Coding Ba Zai Iya Zama Tsoho A Harkar Banki Ba

Yana da kyau a faɗi takamaiman dalilin da ya sa vibe coding — ɗan gajeren umarni, duba sakamako, sake gwadawa — yake kasawa a matsayin tsarin aiki na tsoho a cikin harkar da ke ƙarƙashin ƙa'idodi. Yanayin kasawar ba wai wanda yake a fili ba ne (cewa LLM wani lokaci tana yin barna/hallucinate). Yanayin kasawar na tsari ne kuma yana fitowa a wurare huɗu lokaci guda.

Na farko shi ne rashin yarjejeniyar da aka raba (lack of shared conventions). Injiniyoyi da yawa da ke aiki ta hanyar tattaunawa za su samar da hanyoyi biyar daban-daban don yin abu ɗaya a cikin lamba ɗaya a cikin zangon shekara guda. A cikin yanayin da ba shi da ƙa'idodi, wannan bashi ne na fasaha (technical debt). A cikin yanayin da ke da ƙa'idodi, wannan ita ce hanyar da ke lalacewa lokacin bincike.

Na biyu shi ne lalacewar mahallin (context decay). Wakilan AI ba su da tabbas (stateless). A kan babban aiki, tattaunawa tana wuce iyakokin mahallin (context windows), kuma dalilan da ke tattare da yanke shawara na tsarin gine-gine na farko suna baje. Irin wannan wakilin, bayan makonni biyu, zai yanke shawarar akasin haka a cikin sabuwar tattaunawa saboda babu abin da ke adana dalilin farko. Ga tsarin da ke buƙatar hanyar bincike (audit trail) don masu sarrafawa, wannan bai dace da tsari ba.

Na uku shi ne taruwar lahani na shiru (invisible defect accumulation). Sakamakon binciken Pearce, Yan, da SonarQube da aka ambata a sama ba batutuwa ne na gefe ba. Su ne ƙimar tsohuwa da LLMs ke samar da lamba mai rauni idan babu tsarin bayanai da kuma gwaji mai tsanani. Bankin da ke gudanar da vibe-coding a cikin samarwa yana tara waɗannan lahanoni a kan ƙimar ɗaya, ba tare da samun damar sanin ainihin abin da aka tura ba.

Na huɗu shi ne matsalar gano ƙa'ida (regulatory traceability problem). Sashe na 12 na EU AI Act yana buƙatar shigar da bayanan shigarwa da fitarwa ta atomatik (automatic logging) don tsarin AI masu haɗari sosai. SR 11-7 yana buƙatar rubutattun ayyuka na mai model da mai tantancewa (validator), sarrafa canje-canje (change management) don sabunta model, da kuma ba da rahoto ga hukumar gudanarwa akan haɗarin model na AI. DORA tana buƙatar cikakken sarrafa haɗarin ICT tare da hujjoji da aka rubuta. Babu ɗaya daga cikin waɗannan wajibai da za a iya cika su ta hanyar aikin da babban kayansa shi ne tarihin tattaunawa wanda babu wanda ke adana shi.

Kammalawa ba wai cewa LLMs ba su dace da harkar banki ba ne. Kammalawa ita ce cewa tsarin aiki da ke kewaye da su dole ne ya samar da bayanai na musamman (specifications), hanyoyin bincike (audit trails), da kuma ƙofofin tantancewa (verification gates) a matsayin kayan aiki na farko maimakon a matsayin abin da ake tunani dina baya. Wannan shi ne ainihin abin da spec-driven development yake a aikace.

Spec-Driven Development A Cikin Gidajen Da Ke Karƙashin Ƙa'idodi

Spec-driven development (SDD) yana juya tsarin aiki. Maimakon tsallakewa zuwa aiwatarwa da kuma sake gwadawa tare da wakili, ƙungiyar tana samar da bayanan ƙayyadaddun bayanai (specification) tukunna — yanke shawara na tsarin gine-gine, buƙatu, kwangilolin haɗin gwiwa (interface contracts), ƙa'idodin nasara, sharuɗɗan tsaro — kuma wakilin yana samar da lambar da ke cika ƙayyadaddun bayanan. Tabbatarwa na da tsari: spec yana bayyana abin da sakamakon dole ne ya yi, kuma wani tsari na daban (samar da gwaji, nazarin lamba, tabbatarwa ta hukuma inda ya dace) yana duba ko an yi hakan.

Kayan aiki na aikace-aikace sun haɗu a ƙarshen 2025 da farkon 2026. GitHub Spec Kit ⧉ (wanda aka fitar a ƙarshen 2025) yana tsara niyya kafin samar da lamba. AWS tana haɗa tsarin aiki na spec-first kai tsaye a cikin Kiro IDE nata. JetBrains da Cursor sun gabatar da hanyoyin tsara abubuwa waɗanda ke tsara hulɗar AI. Tsarin aiki kamar BMAD (Breakthrough Method for Agile AI-Driven Development) suna tura gaba tare da ƙungiyoyin wakilan AI na musamman waɗanda ke nuna matsayin mai nazari (analyst), mai zane-zane (architect), mai haɓakawa (developer), da kuma QA a duk faɗin SDLC. Constitutional SDD, wanda aka tsara a cikin takarda arXiv a watan Fabrairu na 2026, yana shigar da takamaiman sharuɗɗan tsaro tare da taswirar rauni na CWE a cikin ƙayyadaddun bayanan kanta.

Ga banki, nau'in da ke da mahimmanci shi ne abin da binciken Augment Code ya kira spec-anchored development — ƙayyadaddun bayanai suna zuwa tukunna, AI yana samar da lambar da aka taƙaita ta hanyarsu, kuma ƙarin matakan gudanarwa (constitutional constraints, supervision checkpoints, human approval gates) suna zama tsakanin samarwa da haɗawa (merge). Wannan shi ne kawai nau'in da ke samar da audit trail wanda Sashe na 12 na EU AI Act ke tsammani, aikin validator da aka rubuta wanda SR 11-7 ke buƙata, da kuma tsarin change-management da DORA ke buƙata.

Zuba jari da ake buƙata na gaske ne, amma kuma ana iya sarrafa shi. Cibiyoyin da ke yin wannan da kyau sun mayar da aikin yau da kullum na injiniyoyi daga rubuta haruffa zuwa samar da abubuwa guda biyu: spec wanda wakilin zai cika, da kuma hanyar tabbatarwa (verification harness) wadda sakamakon dole ne ya wuce. Bukatar tunani akan injiniya ta fi girma a wasu fannoni (bayyana niyya yana da mahimmanci fiye da kowane lokaci) kuma ta ragu a wasu (aikin injiniya na rubuta boilerplate ya ƙare). Cibiyoyin da ba su riga sun yi wannan canjin ba har yanzu suna aiki a cikin yanayin da LLM take zama mai rubutu mafi sauri kawai. Wannan matsayi ba zai iya wanzuwa ba a cikin harka mai ƙa'idodi bayan watanni goma sha biyu masu zuwa.

Tarihin Ƙa'idodi Da Ke Aiki Yanzu

Iyakokin ƙa'idodin 2026 game da AI a cikin harkar banki ba jerin abubuwan dubawa (checklist) ba ne kawai yanzu; tari ne na wajibai da suka haɗu waɗanda ke buƙatar a yi tunaninsu tare. Rana mafi mahimmanci ita ce 2 ga Agusta 2026, lokacin da wajibai na tsarin haɗari sosai na EU AI Act za su fara aiki gaba ɗaya ⧉. Annex III a sarari ya sanya credit scoring, creditworthiness assessment, kimantawa na haɗari a cikin inshorar rayuwa da lafiya, da kuma kimantawa ko rarrabuwar matsayin kuɗi na mutane a matsayin babban haɗari (high-risk). Wajibai da ke fitowa daga wannan rarrabuwa sun haɗa da conformity assessments, tsarin gudanar da inganci (quality management systems), tsarin gudanar da haɗari (risk management frameworks), takaddun fasaha (technical documentation), rajista a cikin bayanan EU (EU database registration), gudanar da bayanai mai ƙarfi (robust data governance), kulawar ɗan adam, da kuma kariya ta intanet (cybersecurity protections). Penalties don keta wajibai masu babban haɗari ya kai Yuro miliyan 35 ko kashi 7% na kuɗin shiga na shekara-shekara na duniya, duk wanda ya fi girma.

Kusa da AI Act ɗin akwai:

DORA (Digital Operational Resilience Act) yana aiki tun daga Janairu na 2025 kuma yana ƙirƙirar wajibai 22 na sarrafa haɗarin ICT waɗanda a sarari ke rufe tsarin AI da ake amfani da su a cikin mahimman ayyukan kuɗi. Sarrafa haɗarin ICT, sa ido ga ɓangare na uku, rahoton aukuwa (incident reporting), da kuma gwajin juriya na aiki (operational resilience testing) suna aiki ga sassan AI kamar yadda suke aiki ga kowane kayan ICT na daban.
SR 11-7 — jagorar sarrafa haɗarin model na Babban Bankin Amurka (Federal Reserve) da OCC, wanda aka rubuta tun asali a cikin 2011 — an tsawaita shi a cikin aikin mai sarrafawa ⧉ don rufe LLMs da tsarin agentic. Jarrabawar FFIEC yanzu a sarari ta haɗa da ikon gudanar da AI. Cibiyoyin da ba za su iya samar da AI model inventory da kimanta haɗari akan buƙata ba suna karɓar bincike mara kyau a yau.
NIST AI RMF (1.0, Janairu 2023) na radin kai ne a Amurka amma masu sarrafa tarayya suna amfani da shi a matsayin ginshiƙi. Ayyukansa guda huɗu (Govern, Map, Measure, Manage) suna daidaitawa da kyau tare da buƙatun tsari na EU AI Act.
ISO/IEC 42001 (wanda aka buga a Disamba 2023) shi ne ma'auni na farko na AI Management System da za a iya ba da shaida a kansa, an tsara shi kamar ISO/IEC 27001 don tsaron bayanai. Bukatar takardar shaida ta 42001 ta haɓaka sosai a cikin Q1 2026 yayin da buƙatun saye suka fara ambaton sa.
UK SM&CR da Consumer Duty — Tsarin Manyan Manajoji da Takaddun Shaida (SM&CR) yanzu yana buƙatar takamaiman alhakin kowane tsarin AI mai babban haɗari. An tsawaita Consumer Duty a cikin jagorar FCA ta kwanan nan ⧉ don rufe sakamakon abokin ciniki na kasuwanci da AI ya yi tasiri a kai.
Taswirar hanya ta G7 post-quantum (Janairu 2026), tsarin ƙaura mai matakai uku na NCSC, da kuma binciken Project Leap na BIS suna zama tare da wannan tari. Wurin haɗuwa yana da mahimmanci: tsarin AI da aka horar ko aka yi amfani da shi akan tsarin cryptographic wanda ba zai iya tsira da canjin post-quantum ba, tsarin AI ne wanda audit trail dinsa da iƙirarin amincinsa suna da rabin rayuwar ƙasa da shekaru goma. Tattaunawa game da agentic engineering da tattaunawa game da post-quantum cryptography suna ƙara zama tattaunawa ɗaya. (Don ƙarin bayani a ɓangaren cryptographic, duba rubutun Mayu 2026 akan ƙaurawar post-quantum a cikin kuɗin kamfanoni.)

Kwatanta Hanyoyi Guda Uku Na Haɓaka Tare Da Taimakon AI

Fanni	Vibe Coding	Spec-Driven Development	Agentic Engineering
Babban bayani na shiga	Dan gajeren umarni	Ƙayyadaddun bayanai na hukuma	Ƙayyadaddun bayanai + shirin tsara wakilai
Aikin injiniya	Mai maimaita umarni	Mawallafin ƙayyadaddun bayanai	Mai tsara aiki da tabbatarwa
Tsarin sakamako	Samar da lamba kai tsaye	Lamba da aka taƙaita ta spec	Tsarin aiki na wakilai da yawa da ke samar da lamba, gwaje-gwaje, da takaddun bayanai
Audit trail	Tarihin tattaunawa (ba a adana shi ba)	Spec + lambar da aka samar + gwaje-gwaje	Spec + alamomin wakili + abubuwan tabbatarwa
Adadin lahani (LLM-only)	Kashi 10-40% na ƙimar rauni (binciken asali)	An rage shi sosai ta hanyar takunkumin spec	Mafi ƙanƙanta tare da ƙofofin tabbatarwa
Binciken ƙa'ida	Bai isa ba don tsarin AI mai babban haɗari	Ya dace da Sashe na 12 na EU AI Act	An tsara shi don Sashe na 12 + SR 11-7 + DORA
Ya dace da harkar banki?	A'a, don samarwa	E, tare da gudanarwa	E, tare da ƙwararren gudanarwa
Iyakacin ƙarfi	Yana da iyaka ta hanyar umarni na lokaci ɗaya	Yana da iyaka ta kyawun spec	Yana da iyaka ta kyawun tsara aiki

Hanyar bayani: Haɗakar sharhin Karpathy (2026), Binciken Augment Code SDD ⧉, Binciken CGI Spec-Driven Development ⧉, da kuma wallafe-wallafen ilimi akan ƙimar raunin samar da lamba na LLM (Pearce et al., Yan et al., Fu et al., 2023-2025).

Gina Bankin Agentic: Dubi Na Gine-Gine

Matsayi na dabara a bayan waɗannan tsarin aikin shi ne abin da C-suite ke buƙatar mallaka a sarari. Agentic engineering a cikin harkar banki ba shiri ne na yawan aikin mai haɓakawa (developer-productivity) kawai ba. Iko ne na cibiya wanda ke taɓa tafiye-tafiyen abokin ciniki na gaba-ɗaya, cikakken yanayin rayuwar ma'amala, da kuma tsarin cryptographic da audit substrate da ke ƙarƙashin duka biyun. Matakai huɗu na wannan ƙarfin suna buƙatar kulawar gudanarwa kai tsaye, dina sama zuwa ƙasa:

Layer 4 — Agent Control Plane Gudanarwa, audit, kill switches, gano halayen da ba na al'ada ba (behavioural anomaly detection), ikon soke na ɗan adam (human override). Tsarin sa ido na HITL da HOTL a kowane rukunin wakili.

Layer 3 — Agentic Workflows Tafiye-tafiyen abokin ciniki, ayyukan cikin gida, bututun ci gaba (development pipeline). Spec-driven ta tsohuwa don tafiye-tafiye masu babban haɗari.

Layer 2 — Data & Model Layer AIBOM (AI Bill of Materials), rajistar model (model registry), tsarin dawo da bayanai (retrieval substrate), sarrafa sigar prompt-template, tarihin fine-tune.

Layer 1 — Quantum-Safe Foundation ML-KEM, ML-DSA, hybrid PKI, crypto-agility. Tushen da duk wani iƙirarin amincin matakan da ke sama ya dogara da shi.

Layer 1 — Gidauniyar Quantum-Safe (The Quantum-Safe Foundation). Kowane mataki na sama yana ɗauka cewa tsarin cryptographic yana da aminci. Tare da taswirar hanya ta G7, shirin matakai uku na NCSC, da kuma binciken BIS Project Leap duk a bayyane ga jama'a, wannan ba batun da ya shafi wani rukunin mutane kaɗai ba ne yanzu. Tsarin agentic wanda aka sanya wa hannu kan hanyoyin bincikensa (audit trails) ƙarƙashin ECDSA na gargajiya, ko kuma wanda kafa maɓallinsa ya dogara da RSA ko ECDH, za su ga iƙirarin amincin shaidarsu ya ƙare tare da cryptography. Cibiyoyin da suka sami wannan daidai suna tura aikin post-quantum gaba kuma suna ɗaukar ML-KEM, ML-DSA, da hybrid PKI a matsayin substrate wanda kowane mataki na sama na audit da tabbacin aminci ya dogara da shi.

Layer 2 — Matakin Bayanai Da Model (The Data and Model Layer). Wannan shi ne inda AI Bill of Materials (AIBOM) ke zama. Kamar Cryptographic Bill of Materials da ake amfani da shi a cikin tsara ƙaurawar post-quantum, AIBOM ita ce rajistar kowane model, rukunin bayanai (dataset), prompt template, retrieval index, fine-tune, da kuma dogaro na AI na ɓangare na uku wanda cibiyar ke gudanarwa. Shi ne abu na zahiri wanda Sashe na 49 na EU AI Act ke buƙata a aikace, rajistar da jarrabawar SR 11-7 ke buƙata yanzu, da kuma tushen kowane tsarin gudanarwa mai inganci. Yawancin cibiyoyi ba su da shi. Za su buƙaci ɗaya nan da watan Agusta.

Layer 3 — Tsarin Aiki Na Agentic (Agentic Workflows). Wannan shi ne matakin da mafi yawan cibiyoyi ke gina shi a halin yanzu, galibi ba tare da isasshen hankali ga matakan 1, 2, da 4 ba. Tsarin aikin da kansu sun kama tun daga na cikin gida (samar da lamba, rubuta takaddun ƙa'idodi, tsarin sabis na abokan ciniki) zuwa na abokan ciniki (copilots na manajojin dangantaka, shigar da abokan ciniki/onboarding, tsarin KYC, sa ido kan ma'amala, inganta FX) zuwa cikakken cin gashin kai (ayyukan baitulmali/treasury operations, wasu ayyukan kasuwanci da gudanar da haɗari inda amincewar mai sarrafawa ta ba da dama). Tsarin horo a wannan matakin shi ne a ɗauke shi a matsayin injiniyan tsarin (systems engineering), ba haɓaka aikace-aikace (application development) ba — tsarin tsara aiki (orchestration patterns), dokokin haɓaka mataki (escalation rules), human-in-the-loop gates, da kuma fitar da audit sune manyan abubuwan damuwa na zane.

Layer 4 — Farantin Kula Da Wakili (The Agent Control Plane). Wannan shi ne abin da Deloitte ta bayyana a matsayin "dakin kula da wakili" ⧉: bincike na lokaci-da-lokaci, shigar da ayyuka (action logging), gano halayen da ba na al'ada ba, kill switches, da kuma kayan aikin human override da ke kewaye da kowane wakili a cikin samarwa. Asasar Step Finance ba wai gazawar AI ba ce ta fuskar fasaha. Gazawar control-plane ce: wakilan suna da izinin da bai kamata su kasance da shi ba, kuma halayen da ba na al'ada ba da ya kamata su tsayar da aikin ba su yi hakan ba. Cibiyoyin da ke gina control plane tukunna — kafin faɗaɗa tura wakilai — sune waɗanda ba za su ga aukuwa irin na Step-Finance a cikin 2027 ba.

Kwatancen da ya dace ga C-suite ba wai "shin muna yin AI fiye da abokan hamayyarmu?" ba ne. Shi ne ko cibiyar tana mallakar duka matakan huɗu, ko kuwa an damƙa mataki ɗaya ko fiye a shiru ga mai siyarwa (vendor) ba tare da ikon kwangila don cika buƙatun takaddun Sashe na 13 na EU AI Act ba. Matsayi na ƙarshe yana kama da lafiya har sai lokacin da mai sarrafawa ya buɗe tambaya.

Kulawar ɗan adam A Aiki: HITL vs HOTL

Bambanci guda ɗaya a cikin Layer 4 wanda masu sarrafawa suka fi mayar da hankali akai a cikin 2026 shi ne tsakanin samfuran sa ido guda biyu. Dukansu nau'ikan kulawar ɗan adam ne; sun bambanta a cikin jinkiri (latency), sassa (scale), da kuma zato da mai sarrafawa ke son bayarwa game da halayen wakili.

Human-in-the-Loop (HITL) shi ne samfurin da wakili ba zai iya aiwatar da wani aiki mai tasiri ba tare da amincewar ɗan adam ta musamman ba. Wakilin yana shirya shawarar, gabatar da ita, kuma ya jira. Wakilin KYC wanda ke yiwa asusu alama don rufewa amma ba zai iya rufe shi ba tare da sa hannun jami'in bin doka (compliance officer) shi ne HITL. Kasuwancin aiki ne: HITL ya fi aminci kuma yana samar da audit trail na Sashe na 14 na fili, amma ba ya dacewa da tsarin aiki na babban rukunin bayanai da jinkiri kaɗan.

Human-on-the-Loop (HOTL) shi ne samfurin da wakili ke aiwatarwa da kansa a cikin sigogi da aka taƙaita (bounded parameters), tare da mutane da ke sa ido kan telemetry a cikin ainihin lokaci kuma suna riƙe da ikon tsayar da wakilin a kowane lokaci. Wakilin sa ido kan zamba a ainihin lokaci wanda ke toshe ma'amala ta atomatik da ke daidai da takamaiman tsarin haɗari, tare da ƙungiyar ayyuka na mutane da ke kallon yawan faɗakarwa (alert volume) da kuma shiga tsakani akan abubuwan da ba na al'ada ba, shi ne HOTL. Kasuwancin akasin haka ne: HOTL yana faɗaɗawa (scales), amma ya dogara da cewa an saita sigogin wakilin daidai kuma gano halayen da ba na al'ada ba yana gano canji kafin cutarwa ta taru.

Sashe na 14 na EU AI Act ba ya bayyana tsakanin HITL ko HOTL ba; yana buƙatar cewa kulawar ɗan adam ta kasance mai ma'ana (meaningful). Ma'anar aiki ita ce kowane wakili mai babban haɗari da banki ke gudanarwa dole ne ya kasance yana da matsayi a sarari, da takaddun shaida akan wane samfurin ne ke aiki, me yasa, da kuma wace ce hanyar haɓaka mataki (escalation pathway) lokacin da wakilin ya haɗu da yanayi a waje da sigoginsa na taƙaitawa. Yawancin bankuna da ke gudanar da gwaji a cikin 2025 ba su da wannan takaddun. Yawancin bankunan da ke gudanar da wakilai a cikin samarwa nan da watan Agusta na 2026 za su buƙaci shi.

Dokar yanke shawara ba ta da rikitarwa. Don ayyuka masu tasiri, ƙanƙantar yawa, da marasa canzawa — ƙaryataccen bashi ga mutum na gaske, rufewar asusu, izinin tura kuɗi mai yawa (large-value wire authorisation), ƙaddamar da rahoton ƙa'ida — HITL shi ne mafi kyawun tsoho. Don ayyuka na babban rukunin bayanai, masu iya canzawa, da aka taƙaita da sigogi — faɗakarwar sa ido kan ma'amala, rarraba takaddun bayanai, tsarin sabis na abokan ciniki na yau da kullum — HOTL ya dace, muddin tsarin gano halayen da ba na al'ada ba da kuma abubuwan kill-switch sun ƙware. Bankunan da ke ɗaukar kowane tsarin aiki a matsayin HITL ba za su sami moriyar aiki na tsarin agentic ba. Bankunan da ke ɗaukar kowane tsarin aiki a matsayin HOTL a ƙarshe za su sami lokacin Step Finance.

Saye vs Gina: Matsalar Wakilin Ɓangare Na Uku (The Third-Party Agent Problem)

Ainihin gaskiyar 2026 da ta kama yawancin bankuna ita ce ba wai za su gina ƙarfin agentic ba ne da farko. Saye za su yi. Yanayin masu siyarwa (vendors) — dandamalin banki na agentic na Oracle da aka ƙaddamar a watan Fabrairu na 2026, Watsonx na IBM, Copilot suite na Microsoft, AWS Bedrock Agents, Salesforce Agentforce, NowAssist na ServiceNow, da kuma rukunin masu siyar da wakilai na musamman na fintech — yana motsawa da sauri fiye da yadda injiniyan cikin gida na banki zai iya. Sakamakon dabara shi ne cewa mafi yawan wakilan da ke aiki a cikin banki a 2027 za su kasance waɗanda wani ya rubuta su, kuma tambayar gudanarwa ba ita ce "za mu iya amincewa da wakilanmu?" ba kuma, a'a, ita ce "za mu iya amincewa da wakilan da muka saya, kuma za mu iya tabbatarwa mai sarrafawa cewa za mu iya?"

Wannan shi ne babban ƙalubalen da ba a cika lura da shi ba ƙarƙashin DORA. Sashe na 28-30 na ƙa'idar suna sanya sarrafa haɗarin ɓangare na uku na ICT (TPRM) ya zama yanki mai sa ido, tare da buƙatu na sarari da suka shafi tanadin kwangila, sa ido na ci gaba, kimanta haɗarin tattarawa (concentration-risk assessment), da kuma dabarun fita (exit strategies). Hukumar Sa Ido ta Turai tana kula da rajistar masu samar da ICT na ɓangare na uku masu mahimmanci, tare da ikon sa ido kai tsaye akan waɗanda aka ayyana a matsayin haka. Sabuwar ainihin aikin ita ce masu siyar da AI na 2026 — masu samar da model na gaba, masu siyar da dandamalin wakilai, SaaS masu ƙarfin AI — suna ƙara zama ɓangarorin uku na ICT waɗanda aka rubuta DORA don rufe su.

Ga banki a matsayin mai saye, ƙa'idodi na aiki guda uku suna aiki:

Nemi AIBOM daga mai siyarwa. Kowane samfurin wakili da aka saya don amfani a cikin tsarin aiki mai babban haɗari dole ne ya zo tare da takaddun kayan aiki (bill of materials) da aka rubuta wanda ke rufe model na asali, asalin bayanan horo da iyakoki, fine-tunes da aka yi amfani da su, retrieval indices da aka shiga, sigogin prompt-template, da kuma sarkar dogaro ga sassan wakilai na gaba. Wannan shi ne abu na zahiri da banki zai buƙata don cika buƙatun takaddun Sashe na 13 a ƙarƙashin EU AI Act. Bankin ba zai iya samar da shi daga baya ba daga mai siyarwa wanda bai himmatu a cikin kwangilar samar da shi ba.

Gwada akwatin baki (black box), ba littafin bayani (brochure) ba. Kimanta saye na masu siyarwa historically yana mayar da hankali kan kwatanta fasali da kuma tattaunawa da abokan ciniki na ishara. Ga tsarin agentic, wannan bai isa ba. Cibiyar dole ne ta gudanar da gwajin hali (behavioural testing) na wakilin ƙarƙashin yanayi daidai da turawarsa na samarwa da aka yi niyya — ciki har da bincike na garkuwa don allurar umarni (prompt injection), juriya ga injiniyan zamantakewa (hanyar Step Finance), canji a ƙarƙashin sauye-sauyen rarraba bayanai (data-distribution shifts), da kuma jinkiri da yanayin gazawa na kill-switch da hanyoyin override. Yawancin kwangilolin masu siyarwa na yanzu ba su ba da damar wannan zurfin gwajin ba tare da takamaiman shawarwari ba; wannan shawarwarin yana buƙatar faruwa kafin a sanya hannu kan kwangilar, ba bayan haka ba.

Sake tattaunawa akan kwangiloli a ƙarƙashin sharuddan Sashe na 13. Mafi yawan yarjejeniyoyin masu siyar da AI na yanzu ba su haɗa da kowane ɗayan takaddun bayanai, haƙƙoƙin audit, sanarwar canjin model, rahoton aukuwa, ko buƙatun bayyana sub-processor waɗanda EU AI Act da DORA ke nema tare ba. Binciken Regulativ na kamfanonin Burtaniya ⧉ ya bayyana a sarari akan wannan batu: sake nazarin shari'a na yarjejeniyar masu siyarwa yana ɗaukar makonni, kuma mafi yawan cibiyoyi ba za su iya cika Sashe na 13 ba don model ɗin da mai siyar da su bai taɓa kasancewa da alƙawarin shari'a na bayyana ayyukansa na ciki ba. Alƙawarin ƙa'ida yana kan mai tura tsarin (deployer), ba mai siyarwa ba. Ƙungiyoyin saye suna buƙatar sanin hakan kafin sake zagayowar sabuntawa na gaba, ba bayan binciken mai sarrafawa ba.

Abin Da Wannan Ke Nufi Ga Kowane Nau'in Banki

Martani mai daidai ya bambanta. Tsarin da ke ƙasa rarrabuwa ne na kusa, ba umarni ba ne.

Bankunan Matakin Farko Na Duniya (Tier-One Universal Banks)

Cibiyoyin da ke da balance sheets na dala triliyan 1+ da kuma kasancewar duniya sune a lokaci guda suka fi fuskantar haɗari (mafi faɗin iyakar ƙa'idodi, mafi girman tsohon tsarin gado, mafi girman darajar maƙasudi don abokan gaba masu cin gashin kansu) kuma sune mafi kyawun albarkatu. Fifiko na dabara shi ne gina farantin kula (control plane) tukunna — Layer 4 na tsarin gine-gine da ke sama — da kuma kawo tsarin spec-driven development a cikin sashin injiniya na cikin gida kafin faɗaɗa tura wakilai gaba. Sakamakon gasa na samun wannan daidai yana da mahimmanci; sakamakon samun kuskure yana da alaƙa da wanzuwa (existential), idan aka yi la'akari da fuskantar hukunci a ƙarƙashin EU AI Act da kuma fuskantar aiki ga tsarin barazana na matakin GTG-1002.

Bankunan Matsakaici Da Na Yanki (Mid-Tier and Regional Banks)

Tambayar gasa ga bankunan mataki na biyu ta fi ta mataki na farko ƙarfi. Suna fuskantar iyakar ƙa'idodi iri ɗaya ba tare da kasafin kuɗi na gudanarwa iri ɗaya ba, fuskar barazana iri ɗaya ba tare da albarkatun tsaro iri ɗaya ba, da kuma tushen abokin ciniki wanda ke ƙara kwatanta su da fintechs na asalin AI. Amsar aiki ita ce a daidaita sosai akan ƙanƙanin rukunin masu siyarwa da aka tantance (tare da kwangilolin da ke cika buƙatun takaddun Sashe na 13), don saka hannun jari a cikin tsarin spec-driven development maimakon injiniyan dandamali na al'ada (custom platform engineering), da kuma amfani da kayan aikin agentic don taƙaita lokacin sabunta COBOL wanda ya kasance anchor na dabara na shekaru da yawa. Cibiyoyin da suka motsa da wuri a nan za su rufe, sosai, gibi na fasaha tare da bankunan mataki na farko a karon farko a cikin tsaran mutane.

Fintechs, PSPs, Da Cibiyoyin Da Ke Kusa Da Crypto

Rukunin fintech da cibiyoyin biyan kuɗi suna da matsala akasin haka: saurin aiki (agility) yana da girma, gudanarwa galibi yana ƙasa da na sauran bankuna, kuma fuskantar hukunci na EU AI Act, ga matsakaicin fintech, yana da alaƙa da wanzuwa. Tsarin dabara shi ne a bi da AI governance a matsayin ƙofar shirye-shiryen samfur (product-readiness gate) maimakon a matsayin mai rufewa na biyayya kawai — gina AIBOM, audit substrate, da kuma tsarin aiki na spec-driven a cikin al'adun injiniya tun daga farko maimakon gyara su dina baya ƙarƙashin matsin lamba na ƙa'idodi. Ga cibiyoyin da kayan aikin biyan kuɗinsu ke haɗuwa da lokacin ƙarshe na adireshin tsari na SWIFT CBPR+ na watan Nuwamba na 2026, saka hannun jari na agentic-engineering shi ne kuma tsarin halitta don mayar da aikin gyaran adireshin tsari ya zama na masana'antu — dokokin tabbatarwa (validation rules), tilasta ingancin bayanai (data-quality enforcement), da kuma haɗin gwiwar CI-pipeline sune ainihin tsarin da tsarin aiki na spec-driven ke sa su zama masu sauƙin sarrafa.

Sashen Injiniya Na Cikin Gida (Internal Engineering Functions)

Ga injiniyoyi da masu bincike da ke karanta wannan, tsarin aikin da ke da mahimmanci shi ne na yau da kullum. Mayar da tsakiyar aiki daga rubuta haruffa zuwa samar da ƙayyadaddun bayanai (specifications) da hanyoyin tabbatarwa (verification harnesses). Bi da alamomin wakili (agent traces), tsare-tsare na tsakiya, da kuma ƙofofin amincewa a matsayin manyan abubuwa na musamman a cikin sarrafa sigar ku (version control). Saka hannun jari a kayan aiki — Spec Kit, Kiro, hanyar tsara aiki ta Cursor, Claude Code tare da takaddun ƙwarewa na matakin aiki (project-level skill files) — wanda ke sanya ƙayyadaddun bayanai ya zama abu mai dorewa kuma lambar da aka samar ta zama wadda za a iya jefarwa bayan amfani. Canjin ergonomic na gaske ne. Sakamakon ƙwararru shi ne cewa tsarin da ake ɗauka a gaba shi ne kuma tsarin da ke tsira daga binciken ƙa'idodi.

Shirin Aiki Na Makonni 12 Zuwa Agusta 2026

Ga mai ɗaukar nauyin gudanarwa (executive sponsor) wanda ke gudanar da shirin agentic-engineering tsakanin yanzu da ranar aiwatar da EU AI Act, aikin yana matsewa zuwa jerin makonni goma sha biyu. Shirin da ke ƙasa ba cikakke ba ne; shi ne mafi ƙanƙanci da hukumar gudanarwa ya kamata ta tsammaci shirin da ke da aminci ya kammala a ranar 2 ga Agusta 2026.

Makonni 1-2 — Samar da AIBOM. Kafa rajista ta tsakiya (centralised inventory) na kowane tsarin AI, model, rukunin bayanai, prompt template, retrieval index, fine-tune, da kuma dogaro na AI na ɓangare na uku a cikin samarwa ko ƙarƙashin ci gaba. Tuba kowane shigarwa zuwa rarrabuwa na Annex III na EU AI Act. Sakamakon shi ne tushen gaskiya ɗaya wanda CRO, CCO, CISO, da CTO kowannensu zai iya tambaya.

Makonni 3-4 — Rarraba samfurin sa ido a kowane tsarin. Don kowane wakili mai babban haɗari da tasiri, rubuta a sarari ko samfurin sa ido shi ne HITL ko HOTL, dalili, hanyar haɓaka mataki (escalation pathway), da kuma sunan mutumin da ke da alhakin a ƙarƙashin SM&CR (UK) ko makamancin tsarin ƙasa. Inda amsar ba ta fito fili ba, yi amfani da HITL a matsayin tsoho har sai an kammala nazarin.

Makonni 5-6 — Gina ko ƙarfafa Agent Control Plane. Shigar da ayyuka a ainihin lokaci (real-time action logging), gano halayen da ba na al'ada ba, kill-switch da hanyoyin override suna aiki akan kowane wakili na samarwa. Inda control plane bai riga ya kasance ga wani tsarin ba, wannan tsarin yana tafiya cikin matsayin taƙaitaccen tura aiki (restricted-deployment) har sai ya kasance.

Makonni 7-8 — Nazarin kwangilar mai siyarwa. Sashin shari'a da na saye suna duba kowace kwangilar mai siyar da AI mai aiki don haƙƙoƙin takaddun Sashe na 13, sanarwar canjin model, rahoton aukuwa, haƙƙoƙin audit, da kuma bayyana sub-processor. Sakamakon shi ne jeri na matakai: compliant (mai bin doka), remediation required (yana buƙatar gyara), replacement required (yana buƙatar sauyawa). Shawarar sauyawa tana buƙatar farawa yanzu don samun damar kammalawa a wannan shekarar.

Makonni 9-10 — Gwajin gwagwarmaya na conformity assessment. Ga kowane tsarin babban haɗari a ƙarƙashin Annex III, kammala tsarin aiki na conformity-assessment kamar wata hukuma mai ba da shaida za ta zo a mako mai zuwa. Wannan zai fitar da gibin da ke kama da ƙanƙani a kan takarda amma suna da tsanani a cikin aiki yayin bincike. Gyara abin da za a iya gyarawa; rubuta sauran abubuwan.

Makonni 11-12 — Tabbatarwa kafin canji da amincewar hukumar gudanarwa. Nazarin ƙarshe na AIBOM, rarrabuwa na HITL/HOTL, shaidar control-plane, matsayin gyaran mai siyarwa, da kuma sakamakon conformity-assessment. An tabbatar da alhakin babban manaja da aka ambata suna. Hukumar gudanarwa ta rubuta matsayi a mintuna. Sanar da mai sarrafawa inda tsarin ke tsammanin sanarwar rigakafi.

Cibiyoyin da suka kammala wannan jerin makonni goma sha biyu ba wai sun warware agentic engineering gaba ɗaya ba ne. Za su kafa ginshiƙi wanda shirin da ke da aminci yake buƙata. Cibiyoyin da ba su fara ba a lokatim da aka buga wannan rubutun ba su ne kawai masu sakaci ba, kamar yadda binciken Regulativ ya faɗi a ɓangaren SWIFT. Su ne mafi rinjaye. Tambayar da kowane CCO, CRO, da CTO ke buƙatar amsawa a cikin makonni biyu masu zuwa ita ce ko kamfanin zai yi aiki a watan Mayu ko kuma zai yi ruburubu a watan Yuli.

Conclusion

Baban abin lura mai wuya wanda ya fito fili a duk faɗin masana'antar a cikin watanni shida da suka gabata shi ne cewa tsofaffin hanyoyin aiki a babban matakin kamfani ba sabuwar fasaha ce ke wuce su ba amma sabon tsarin aiki ne. Kayan aikin agentic sun bayyana — wani lokaci a cikin samarwa, wani lokaci a cikin rahoton aukuwa — kura-kurai da gibi a cikin tsoffin tsarin gado waɗanda ke taruwa a shiru tsakanin shekaru da yawa. Kayan aiki guda ɗya sun samar da albarkatu ga masu aikata laifuka waɗanda a baya suke buƙatar goyon bayan gwanti. Irin waɗannan kayan aikin, idan aka yi amfani da su a cikin gida tare da horo, sune mafi kyawun hanyar da cibiyoyi ke da ita don rufe gibi na gado, cika lokacin ƙarshe na ƙa'ida na watan Agusta 2026, da kuma kaiwa ga saurin aiki wanda tsammanin abokan ciniki da ainihin gasa ke buƙata yanzu.

Cibiyoyin da suka mallaki wannan matsayi a cikin gida — waɗanda ke ɗaukar agentic engineering a matsayin tsarin tsari na banki maimakon inganta yawan aiki kawai da aka saya daga mai siyarwa — za su shafe shekaru biyu masu zuwa suna tara fa'ida. Cibiyoyin da ba su yi hakan ba za su shafe shekaru biyu masu zuwa suna gano, a cikin rahoton aukuwa da binciken mai sarrafawa, abin da ya kamata su gina. Zaɓi tsakanin waɗannan sakamako guda biyu shawara ce ta matakin hukumar gudanarwa ta 2026, ba shawara ce ta fasaha ta 2028 ba.

Don mahallin baya a kan wannan rukunin yanar gizon, rubutun Afrilu 2026 akan quantum thresholds ya rufe hanyar hardware da ke goyon bayan Layer 1 na tsarin gine-gine da ke sama, rubutun Mayu 2026 akan ƙaurawar post-quantum don kuɗin kamfanoni ya rufe tsarin cryptographic dalla-dalla, nazarin Mayu 2026 na lokacin ƙarshe na structured-address na pacs.008 ya rufe tsarin ƙa'ida da injiniya wanda spec-driven validation ke sa shi ya kasance mai sauƙin sarrafa, kuma ayyukan buɗaɗɗen lamba (open-source) na Rust akan KyberLib, pain001, da pacs008 suna cikin babban ƙoƙarin sanya production-grade primitives — quantum-safe, masu bin ƙa'idodin biyan kuɗi, audit-ready — a hannun ƙungiyoyin injiniya waɗanda za su gina bankin agentic. Haɗin kai tsakanin waɗannan rubuce-rubuce ba haɗari ba ne. Shi ne siffar aikin da shekaru biyu masu zuwa ke buƙata.

Tambayoyi Da Aka Fi Yi (Frequently Asked Questions)

Menene bambanci tsakanin generative AI, agentic AI, da kuma agentic engineering?

Generative AI yana samar da abun ciki a cikin martani ga umarni (prompt); yana da martani (reactive). Agentic AI yana bin manufofin da aka ayyana da kansa, yana shiga bayanai, yin amfani da kayan aiki, da kuma ɗaukar ayyuka a duk faɗin tsarin aiki mai matakai da yawa ba tare da buƙatar umarnin ɗan adam a kowane mataki ba. Agentic engineering — kalmar da Karpathy ya ɗauka a 2026 ⧉ — ita ce tsarin aiki na tsara wakilai akan cikakkun bayanai (specifications) tare da kulawar ɗan adam. Ga harkar banki, bambancin yana da mahimmanci saboda iyakar ƙa'idodi, samfurin barazana, da kuma tsarin injiniya sun bambanta ga kowane nau'i. Hanyar tattaunawa (chat interface) da kuma wakilin kasuwanci mai zaman kansa gaba ɗaya ba sa cikin aji ɗaya na ƙa'ida, kuma bi da su kamar suna aji ɗaya yana haifar da rauni a ɓangarorin biyu.

Me yasa lokacin ƙarshe na EU AI Act na Agusta 2026 deadlines yake da mahimmanci ga bankuna?

Annex III na AI Act a sarari ya sanya wasu manyan sharuɗɗan amfani da AI na banki a matsayin babban haɗari: creditworthiness assessment da credit scoring na mutane na gaske, kimanta haɗari da farashi a cikin inshorar rayuwa da lafiya, da kuma kimantawa ko rarrabuwar matsayin kuɗi na mutane. Daga 2 ga Agusta 2026, masu tura waɗannan tsarin (deployers) dole ne su nuna bin tsarin gudanar da inganci, tsarin gudanar da haɗari, takaddun fasaha, conformity assessments, rajista a cikin bayanan EU, gudanar da bayanai mai ƙarfi, kulawar ɗan adam, da kuma kariya ta intanet. Sashe na 12 yana buƙatar shigar da bayanan shigarwa da fitarwa ta atomatik (automatic logging). Sashe na 14 yana buƙatar kulawar ɗan adam mai ma'ana (HITL ko HOTL, kamar yadda ya dace da tsarin). Tara don keta doka ya kai Yuro miliyan 35 ko kashi 7% na kuɗin shiga na shekara-shekara na duniya. Aikin cika waɗannan wajibai aikin injiniya ne — ba aikin takaddun shaida ba — kuma shi ne dalili na zahiri da ya sa tsarin spec-driven ya haɓaka ta hanyar Q1 2026.

Menene bambancin aiki tsakanin HITL da HOTL, kuma yaushe ya kamata kowanne ya yi aiki?

HITL (Human-in-the-Loop) yana nufin wakilin ba zai iya aiwatar da ayyuka masu tasiri ba tare da amincewar ɗan adam ta musamman ba. HOTL (Human-on-the-Loop) yana nufin wakilin yana aiwatarwa da kansa a cikin sigogi da aka taƙaita, tare da mutane da ke sa ido kan telemetry kuma suna riƙe da ikon tsayar da shi a kowane lokaci. Sashe na 14 na EU AI Act yana buƙatar cewa sa ido ya kasance mai ma'ana amma ba ya bayyana wane samfuri ba. Dokar yanke shawara ita ce a yi amfani da HITL inda aikin yake da tasiri, ƙanƙantar yawa, da kuma maras canzawa (ƙaryataccen bashi, rufewar asusu, izinin tura kuɗi mai yawa, ƙaddamar da rahoton ƙa'ida); da kuma HOTL inda aikin yake da babban rukunin bayanai, mai iya canzawa, da kuma taƙaitaccen sigogi (faɗakarwar sa ido kan ma'amala, rarraba takaddun bayanai, tsarin sabis na abokan ciniki na yau da kullum). Dukansu biyun suna buƙatar kill-switch da kuma override infrastructure su kasance suna aiki kuma an gwada su; bambancin shi ne ko mutumin yana sama da aiwatarwa (HITL) ko kuwa yana kusa da shi (HOTL).

Mafi yawan wakilanmu za su fito ne daga masu siyarwa. Yaya za mu cika buƙatun DORA da EU AI Act don tsarin da ba mu gina ba?

Alƙawarin ƙa'ida yana kan mai tura tsarin (deployer), ba mai siyarwa (vendor) ba. Amsar aiki tana da fuskoki uku. Na farko, nemi takamaiman AIBOM da aka rubuta daga mai siyarwa kafin sanya hannu — tarihin model, asalin bayanan horo, fine-tunes, prompt templates, retrieval indices, sarkokin dogaro. Na biyu, gudanar da gwajin hali na wakilin ƙarƙashin yanayi daidai da na samarwa, ciki har da bincike na garkuwa don allurar umarni (prompt injection) da juriya ga injiniyan zamantakewa. Na uku, sake tattaunawa akan kwangilolin masu siyarwa don haɗawa da haƙƙoƙin takaddun Sashe na 13, sanarwar canjin model, rahoton aukuwa, haƙƙoƙin audit, da kuma bayyana sub-processor — mafi yawan kwangilolin na yanzu ba su da kowane ɗayan waɗannan. Sashe na 28-30 na DORA yana rufe sarrafa haɗarin ɓangare na uku na ICT kuma sune ginshiƙan ƙa'idodi masu dacewa a ɓangaren Turai; jagorar FFIEC ita ce makamancin haka a ɓangaren Amurka. Aikin yana da mahimman gaske; ba za a iya jinkirta shi ba.

Yaya damuwar da ya kamata bankuna su kasance da ita game da abokan gaba na agentic?

Amsa ta gaskiya ita ce barazanar ta gaske ce kuma ta bambanta ta fuskar aiki da barazanar intanet ta baya. Bayyana GTG-1002 da Anthropic ta yi a watan Nuwamba na 2025 shi ne misali mafi kyau: agentic AI da ke gudanar da kashi 80-90% na ayyukan dabaru a cikin kamfen ɗin leƙen asiri mai samun goyon bayan gwamnati akan kusan maƙasudai talatin na tsaro, makamashi, da fasaha, yana aiki a kan dubu-dubu na buƙatu a kowane daƙiƙa. Aukuwar Step Finance a watan Janairu na 2026 — asasar dala miliyan 27-30 da wakilan kasuwanci na AI suka haifar tare da izini mai yawa — shi ne misali na yadda tura AI na cikin gida zai iya zama fuskar hari (attack surface). Flashpoint 2026 GTIR ta lura da hawan kashi 1,500% a cikin tattaunawa ta haramtacciyar hanya mai alaƙa da AI a cikin wata guda. Waɗannan ba labarun zato ba ne; abubuwa ne na rahoton aukuwa na 2025-2026. Bankunan da ke gudanar da ayyukan tsaro na gargajiya akan abokan gaba na agentic suna fuskantar haɗari marar daidaito (asymmetrically exposed), kuma martani mai kyau shi ne gina ƙarfin tsaro na AI-akan-AI maimakon rage canjin agentic a ɓangaren kai hari.

Shin agentic AI "ChatGPT tare da sabobin MCP" kawai ne?

A'a, kuma wannan yana ɗaya daga cikin mafi rikitarwa kuskuren fahimta a kasuwar yanzu. Hanyar sadarwa ta tattaunawa da aka ƙarfafa da sabobin MCP tsari ne mai amimci don dawo da bayanai da aiki akansu a cikin wani zama da aka taƙaita (bounded session). Agentic engineering tsari ne na cibiya — AIBOM, agent control plane, spec-driven development pipeline, audit substrate, quantum-safe cryptographic foundation, da kuma tsarin tsara aiki a duk faɗin tafiye-tafiyen abokin ciniki na gaba-ɗaya. Waɗannan ba fasaloli ba ne da aka saya daga mai siyarwa; matsayi ne na mallakar cibiya. Bankunan da ke ɗaukar wannan tambayar a matsayin shawarar saye suna ƙarewa da tura tsarin maras zurfi wanda ke kasawa lokacin bincike. Bankunan da ke ɗaukar ta a matsayin tambayar injiniya da mallakar gudanarwa suna ƙarewa da kadara mai ƙaruwa.

Menene abu ɗaya mafi mahimmanci da ya kamata banki ya kasance yana yi a cikin makonni goma sha biyu masu zuwa?

Abubuwa uku, masu biyo juna. Na farko, samar da AI Bill of Materials — cikakken rajista na kowane tsarin AI, model, rukunin bayanai, prompt template, retrieval index, da kuma dogaro na AI na ɓangare na uku a cikin samarwa ko ƙarƙashin ci gaba, tare da kowane shigarwa da aka rarraba ƙarƙashin Annex III na EU AI Act. Cibiyar da ba za ta iya samar da wannan ba lokacin da mai sarrafawa ya nema ita ce cibiyar da za ta karbi bincike mara kyau. Na biyu, gina agent control plane don kowane tsarin AI da ke yanke shawara a halin yanzu ko kuma ke tasiri sosai akan shawarwarin da ke shafar abokan ciniki — audit logging, behavioural anomaly detection, human override, da kuma kill switches a matsayin tsarin tsoho, ba a matsayin abu na gaba ba. Na uku, mayar da al'adun injiniya na cikin gida daga vibe coding zuwa spec-driven development akan aikin da ya fi mahimmanci — tsarin babban haɗari, tsarin aiki mai ƙa'idodi, da kuma bututun sabunta gado. Na farko guda biyu aikin biyayya ne (compliance work); na uku aikin gasa ne (competitive work). Cibiyoyin da ke yin duka ukun za su kasance a cikin matsayi mafi ƙarfi fiye da waɗanda ke yin ɗaya ko babu ko ɗaya. An tsara cikakken jerin makonni goma sha biyu a cikin sashin shirin aiki na sama.

Hanyoyin Bayani (References)

Sebastien Rousseau, (2026). Kare Littafin Kuɗi: Jagorar Hukumar Gudanarwa Ga Hijirar Post-Quantum Don Kuɗin Kamfani.
Sebastien Rousseau, (2026). Lokacin Ƙarshe na pacs.008 Structured-Address na Nuwamba 2026: Dubi na Watanni Shida.
Sebastien Rousseau, (2026). Quantum Thresholds Suna Motsawa Kuma.
Sebastien Rousseau, (2023). CRYSTALS-Kyber: Algorithm na Kariya a cikin Zamanin Quantum.
Mansurova, M. (2026). Daga Vibe Coding zuwa Spec-Driven Development ⧉. Towards Data Science.
CGI, (2026). Spec-driven development: Daga vibe coding zuwa intent engineering ⧉. CGI.
Augment Code, (2026). Mene ne Spec-Driven Development? Cikakken Jagora ⧉. Augment Code.
BCMS, (2026). Spec-Driven Development: Cikakken Jagora na 2026 ⧉. BCMS.
Deloitte, (2026). Sarrafa sabuwar igiyar haɗari daga wakilan AI a cikin banki ⧉. Deloitte Center for Financial Services.
Anthropic, (2025). Gano da kuma magance mummunan amfani da AI: Agusta 2025 ⧉. Anthropic.
Anthropic, (2025). Dakatar da kamfen na farko na leƙen asiri na intanet da AI ya tsara ⧉. Anthropic.
Flashpoint, (2026). Rahoton Bayanan Barazana na Duniya na 2026 ⧉. HSToday / Flashpoint.
Beam AI, (2026). Keta Tsaro Guda 5 Na Gaske Na Wakilin AI a Cikin 2026 da Darussansu ⧉. Beam.
Congressional Research Service, (2026). Agentic Artificial Intelligence da Hare-haren Intanet ⧉. Congress.gov.
European Commission, (2024). Dokar (EU) 2024/1689 akan Hankalin Injin (EU AI Act).
Knowlee, (2026). EU AI Act don Sabis na Kuɗi: Wajibai na Annex III ga Bankuna, Fintechs, da Insurers ⧉. Knowlee.
Regulativ, (2026). Lokacin Ƙarshe na EU AI Act na Agusta 2026: Abin da Kamfanonin Sabis na Kuɗi Dole ne Su Yi Yanzu ⧉. Finextra.
AegisAI Compliance, (2026). AI Governance ga Bankuna: Jagorar Biyayya Ta SR 11-7 & EU AI Act ⧉. AegisAI.
The Financial Brand, (2026). Yadda Wakilan AI Masu Zaman Kansu Za Su Sake Fassaƙa Haɓakar Banki Da Gaske ⧉. The Financial Brand.
UXDA, (2026). Shin Bankin Agentic AI Zai Iya Yi Wa Sabis na Kuɗi na Duniya Barazana? ⧉. UXDA.
Neurons Lab, (2026). Agentic AI a cikin Sabis na Kuɗi: Binciken Bincike na 2026 ⧉. Neurons Lab.
Computer Weekly, (2026). AI don taimaka wa mainframes su kasance masu mahimmanci ga kasuwanci a cikin 2026 ⧉. Computer Weekly.
The Financial Revolutionist, (2026). Tsoffin tsarin gado suna jefa bankuna a cikin haɗari ⧉. The Financial Revolutionist.
CIO Magazine, (2025). Amfani da AI don sabunta mainframes: Maida tsohuwar fasaha ta zama fa'ida ta dabara ⧉. CIO Magazine.
Microsoft Azure, (2025). Yadda Muke Amfani da Wakilan AI don Hijirar COBOL da Sabunta Mainframe ⧉. Microsoft DevBlogs.
VentureBeat, (2026). Agentic coding a matakin kamfani yana buƙatar spec-driven development ⧉. VentureBeat.
The News (Pakistan), (2026). Dalilin da ya sa yunƙurin bankunan Burtaniya na agentic AI ke damun masu sarrafa UK ⧉. The News International.
CNBC, (2026). Mythos na Anthropic ya haifar da 'hysteria' na tsaron intanet ⧉. CNBC.

Wanda aka duba na ƙarshe 2026-06-23.

Bita ta ƙarshe 2026-07-18.

Sake buga wannan labari

Agentic Engineering na Bankuna: Tsarin 2026 na C-Suite da Injiniyoyin da Zasu Gina Shi — Sebastien Rousseau

Agentic AI ya wuce daga gwaji zuwa samarwa a bankunan duniya. Kashi 70% na cibiyoyi suna amfani da shi; ɗaya cikin biyar ne kawai ke da ƙwararre na mulki. Abokan gaba masu cin gashin kansu suna aiki da saurin injin, gado na COBOL an rubuta shi don zato na sarrafawa-tari na 1960s, kuma EU AI Act ta Agusta 2026 tana nesa da makonni goma sha biyu.

An lasisin wannan labari a karkashin Creative Commons Attribution 4.0 International. Sake bugawa na bukatar nuna asalin URL na asali.

Agentic Engineering na Bankuna: Tsarin 2026 na C-Suite da Injiniyoyin da Zasu Gina Shi — Sebastien Rousseau

Agentic AI ya wuce daga gwaji zuwa samarwa a bankunan duniya. Kashi 70% na cibiyoyi suna amfani da shi; ɗaya cikin biyar ne kawai ke da ƙwararre na mulki. Abokan gaba masu cin gashin kansu suna aiki da saurin injin, gado na COBOL an rubuta shi don zato na sarrafawa-tari na 1960s, kuma EU AI Act ta Agusta 2026 tana nesa da makonni goma sha biyu.

Originally published at https://sebastienrousseau.com/ha/2026-05-17-agentic-engineering-banks-blueprint-2026/ by Sebastien Rousseau.
Licensed under CC-BY-4.0.

SEBASTIEN ROUSSEAU FOUNDER · ENGINEER