{"id":127791,"date":"2026-07-01T15:02:38","date_gmt":"2026-07-01T08:02:38","guid":{"rendered":"https:\/\/tino.vn\/blog\/?p=127791"},"modified":"2026-07-01T15:19:43","modified_gmt":"2026-07-01T08:19:43","slug":"lo-trinh-hoc-llms-ai-agents","status":"publish","type":"post","link":"https:\/\/tino.vn\/blog\/lo-trinh-hoc-llms-ai-agents\/","title":{"rendered":"L\u1ed9 tr\u00ecnh h\u1ecdc LLMs & AI Agents: L\u00e0m sao \u0111\u1ec3 b\u1eaft \u0111\u1ea7u \u0111\u00fang h\u01b0\u1edbng?"},"content":{"rendered":"\n

L\u1ed9 tr\u00ecnh h\u1ecdc LLMs & AI Agents \u0111ang tr\u1edf th\u00e0nh ch\u1ee7 \u0111\u1ec1 \u0111\u01b0\u1ee3c nhi\u1ec1u ng\u01b0\u1eddi quan t\u00e2m khi AI kh\u00f4ng c\u00f2n d\u1eebng l\u1ea1i \u1edf vi\u1ec7c h\u1ecfi \u2013 \u0111\u00e1p, m\u00e0 \u0111\u00e3 ti\u1ebfn \u0111\u1ebfn giai \u0111o\u1ea1n c\u00f3 th\u1ec3 l\u1eadp k\u1ebf ho\u1ea1ch, d\u00f9ng c\u00f4ng c\u1ee5, truy xu\u1ea5t d\u1eef li\u1ec7u v\u00e0 h\u1ed7 tr\u1ee3 t\u1ef1 \u0111\u1ed9ng h\u00f3a quy tr\u00ecnh. N\u1ebfu tr\u01b0\u1edbc \u0111\u00e2y ng\u01b0\u1eddi h\u1ecdc ch\u1ec9 c\u1ea7n bi\u1ebft c\u00e1ch \u0111\u1eb7t c\u00e2u l\u1ec7nh cho ChatGPT<\/a>, Gemini hay Claude, th\u00ec hi\u1ec7n t\u1ea1i, k\u1ef9 n\u0103ng quan tr\u1ecdng h\u01a1n l\u00e0 hi\u1ec3u c\u00e1ch m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn ho\u1ea1t \u0111\u1ed9ng, c\u00e1ch k\u1ebft n\u1ed1i d\u1eef li\u1ec7u ri\u00eang, c\u00e1ch x\u00e2y d\u1ef1ng AI Agent v\u00e0 c\u00e1ch \u0111\u00e1nh gi\u00e1 ch\u1ea5t l\u01b0\u1ee3ng \u0111\u1ea7u ra. <\/strong><\/p>\n\n\n\n

C\u00f9ng Tino t\u00ecm hi\u1ec3u chi ti\u1ebft qua b\u00e0i vi\u1ebft d\u01b0\u1edbi \u0111\u00e2y nh\u00e9!<\/strong><\/p>\n\n\n\n

L\u1ed9 tr\u00ecnh h\u1ecdc LLMs & AI Agents l\u00e0 g\u00ec v\u00e0 t\u1ea1i sao quan tr\u1ecdng trong?<\/strong><\/h2>\n\n\n\n

L\u1ed9 tr\u00ecnh h\u1ecdc LLMs & AI Agent l\u00e0 chu\u1ed7i c\u00e1c b\u01b0\u1edbc h\u1ecdc t\u1eadp \u0111\u01b0\u1ee3c s\u1eafp x\u1ebfp theo m\u1ee9c \u0111\u1ed9 kh\u00f3 t\u0103ng d\u1ea7n, gi\u00fap ng\u01b0\u1eddi h\u1ecdc \u0111i t\u1eeb vi\u1ec7c hi\u1ec3u b\u1ea3n ch\u1ea5t m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn (Large Language Models) cho \u0111\u1ebfn kh\u1ea3 n\u0103ng t\u1ef1 thi\u1ebft k\u1ebf, x\u00e2y d\u1ef1ng v\u00e0 v\u1eadn h\u00e0nh m\u1ed9t t\u00e1c nh\u00e2n AI (AI Agent) c\u00f3 th\u1ec3 t\u1ef1 ra quy\u1ebft \u0111\u1ecbnh, s\u1eed d\u1ee5ng c\u00f4ng c\u1ee5 v\u00e0 ho\u00e0n th\u00e0nh nhi\u1ec7m v\u1ee5 m\u00e0 kh\u00f4ng c\u1ea7n con ng\u01b0\u1eddi can thi\u1ec7p li\u00ean t\u1ee5c.<\/p>\n\n\n\n

Theo kh\u1ea3o s\u00e1t g\u1ea7n \u0111\u00e2y v\u1edbi h\u01a1n 500 l\u00e3nh \u0111\u1ea1o k\u1ef9 thu\u1eadt, h\u01a1n m\u1ed9t n\u1eeda s\u1ed1 t\u1ed5 ch\u1ee9c \u0111\u00e3 tri\u1ec3n khai AI Agent cho c\u00e1c quy tr\u00ecnh nhi\u1ec1u giai \u0111o\u1ea1n v\u00e0 ph\u1ea7n l\u1edbn c\u00f3 k\u1ebf ho\u1ea1ch m\u1edf r\u1ed9ng sang nh\u1eefng tr\u01b0\u1eddng h\u1ee3p s\u1eed d\u1ee5ng ph\u1ee9c t\u1ea1p h\u01a1n trong t\u01b0\u01a1ng lai. \u0110\u00e2y ch\u00ednh l\u00e0 l\u00fd do \u0111\u1ecbnh ngh\u0129a v\u1ec1 vai tr\u00f2 k\u1ef9 thu\u1eadt li\u00ean quan \u0111\u1ebfn AI \u0111ang thay \u0111\u1ed5i: nh\u00e0 tuy\u1ec3n d\u1ee5ng kh\u00f4ng ch\u1ec9 t\u00ecm ng\u01b0\u1eddi bi\u1ebft d\u00f9ng ChatGPT, m\u00e0 t\u00ecm ng\u01b0\u1eddi bi\u1ebft x\u00e2y d\u1ef1ng h\u1ec7 th\u1ed1ng Agentic AI v\u00e0 \u1ee9ng d\u1ee5ng LLMs v\u00e0o b\u00e0i to\u00e1n th\u1ef1c t\u1ebf c\u1ee7a doanh nghi\u1ec7p.<\/p>\n\n\n

\n
\"L\u1ed9
L\u1ed9 tr\u00ecnh h\u1ecdc LLMs & AI Agents l\u00e0 g\u00ec v\u00e0 t\u1ea1i sao quan tr\u1ecdng trong?<\/strong><\/figcaption><\/figure>\n<\/div>\n\n\n

\u0110i\u1ec3m kh\u00e1c bi\u1ec7t l\u1edbn nh\u1ea5t gi\u1eefa m\u1ed9t chatbot th\u00f4ng th\u01b0\u1eddng v\u00e0 m\u1ed9t AI Agent n\u1eb1m \u1edf kh\u1ea3 n\u0103ng t\u1ef1 ch\u1ee7. M\u1ed9t chatbot ch\u1ec9 tr\u1ea3 l\u1eddi khi \u0111\u01b0\u1ee3c h\u1ecfi, c\u00f2n m\u1ed9t AI Agent c\u00f3 th\u1ec3 t\u1ef1 l\u00ean k\u1ebf ho\u1ea1ch, g\u1ecdi c\u00f4ng c\u1ee5 b\u00ean ngo\u00e0i, ki\u1ec3m tra k\u1ebft qu\u1ea3 v\u00e0 \u0111i\u1ec1u ch\u1ec9nh h\u00e0nh \u0111\u1ed9ng cho \u0111\u1ebfn khi ho\u00e0n th\u00e0nh m\u1ee5c ti\u00eau \u0111\u01b0\u1ee3c giao. Ch\u00ednh n\u0103ng l\u1ef1c n\u00e0y khi\u1ebfn vi\u1ec7c h\u1ecdc LLMs & AI Agent tr\u1edf th\u00e0nh k\u1ef9 n\u0103ng c\u00f3 gi\u00e1 tr\u1ecb cao tr\u00ean th\u1ecb tr\u01b0\u1eddng lao \u0111\u1ed9ng hi\u1ec7n nay.<\/p>\n\n\n\n

Ki\u1ebfn th\u1ee9c n\u1ec1n t\u1ea3ng c\u1ea7n c\u00f3 tr\u01b0\u1edbc khi b\u01b0\u1edbc v\u00e0o l\u1ed9 tr\u00ecnh h\u1ecdc LLMs & AI Agent<\/strong><\/h2>\n\n\n\n

Tr\u01b0\u1edbc khi \u0111i s\u00e2u v\u00e0o c\u00e1c k\u1ef9 thu\u1eadt n\u00e2ng cao, ng\u01b0\u1eddi h\u1ecdc c\u1ea7n chu\u1ea9n b\u1ecb m\u1ed9t n\u1ec1n t\u1ea3ng v\u1eefng ch\u1eafc. \u0110\u00e2y l\u00e0 b\u01b0\u1edbc kh\u00f4ng th\u1ec3 b\u1ecf qua trong b\u1ea5t k\u1ef3 l\u1ed9 tr\u00ecnh h\u1ecdc LLMs & AI Agent nghi\u00eam t\u00fac n\u00e0o.<\/p>\n\n\n\n

T\u01b0 duy l\u1eadp tr\u00ecnh c\u01a1 b\u1ea3n v\u1edbi Python<\/strong><\/h3>\n\n\n\n

Python hi\u1ec7n l\u00e0 ng\u00f4n ng\u1eef l\u1eadp tr\u00ecnh ph\u1ed5 bi\u1ebfn nh\u1ea5t trong l\u0129nh v\u1ef1c AI nh\u1edd h\u1ec7 sinh th\u00e1i th\u01b0 vi\u1ec7n phong ph\u00fa. Ng\u01b0\u1eddi m\u1edbi n\u00ean b\u1eaft \u0111\u1ea7u v\u1edbi c\u00fa ph\u00e1p c\u01a1 b\u1ea3n, sau \u0111\u00f3 l\u00e0m quen v\u1edbi c\u00e1c th\u01b0 vi\u1ec7n x\u1eed l\u00fd d\u1eef li\u1ec7u nh\u01b0 Pandas v\u00e0 NumPy. Vi\u1ec7c th\u1ef1c h\u00e0nh l\u00e0m s\u1ea1ch m\u1ed9t b\u1ed9 d\u1eef li\u1ec7u th\u1ef1c t\u1ebf s\u1ebd gi\u00fap h\u00ecnh th\u00e0nh t\u01b0 duy l\u00e0m vi\u1ec7c v\u1edbi d\u1eef li\u1ec7u, m\u1ed9t k\u1ef9 n\u0103ng n\u1ec1n t\u1ea3ng cho m\u1ecdi c\u00f4ng vi\u1ec7c li\u00ean quan \u0111\u1ebfn LLMs sau n\u00e0y.<\/p>\n\n\n

\n
\"Ki\u1ebfn
Ki\u1ebfn th\u1ee9c n\u1ec1n t\u1ea3ng c\u1ea7n c\u00f3 tr\u01b0\u1edbc khi b\u01b0\u1edbc v\u00e0o l\u1ed9 tr\u00ecnh h\u1ecdc LLMs & AI Agent<\/strong><\/figcaption><\/figure>\n<\/div>\n\n\n

Hi\u1ec3u v\u1ec1 d\u1eef li\u1ec7u, x\u00e1c su\u1ea5t v\u00e0 to\u00e1n h\u1ecdc \u1ee9ng d\u1ee5ng<\/strong><\/h3>\n\n\n\n

Kh\u00f4ng c\u1ea7n tr\u1edf th\u00e0nh chuy\u00ean gia to\u00e1n h\u1ecdc, nh\u01b0ng ng\u01b0\u1eddi h\u1ecdc n\u00ean n\u1eafm \u0111\u01b0\u1ee3c c\u00e1c kh\u00e1i ni\u1ec7m c\u01a1 b\u1ea3n v\u1ec1 x\u00e1c su\u1ea5t, th\u1ed1ng k\u00ea v\u00e0 \u0111\u1ea1i s\u1ed1 tuy\u1ebfn t\u00ednh. Nh\u1eefng ki\u1ebfn th\u1ee9c n\u00e0y gi\u00fap hi\u1ec3u v\u00ec sao m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef \u0111\u01b0a ra d\u1ef1 \u0111o\u00e1n theo c\u00e1ch x\u00e1c su\u1ea5t, thay v\u00ec t\u1ea1o c\u1ea3m gi\u00e1c m\u00f4 h\u00ecnh \u0111ang “suy ngh\u0129” gi\u1ed1ng con ng\u01b0\u1eddi.<\/p>\n\n\n\n

Giai \u0111o\u1ea1n 1 \u2013 L\u00e0m quen v\u1edbi LLMs: Hi\u1ec3u c\u00e1ch m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn ho\u1ea1t \u0111\u1ed9ng<\/strong><\/h2>\n\n\n\n

\u0110\u00e2y l\u00e0 b\u01b0\u1edbc kh\u1edfi \u0111\u1ea7u ch\u00ednh th\u1ee9c trong l\u1ed9 tr\u00ecnh h\u1ecdc LLMs & AI Agent. M\u1ee5c ti\u00eau c\u1ee7a giai \u0111o\u1ea1n n\u00e0y l\u00e0 hi\u1ec3u b\u1ea3n ch\u1ea5t k\u1ef9 thu\u1eadt \u0111\u1ee9ng sau c\u00e1c m\u00f4 h\u00ecnh nh\u01b0 GPT, Claude hay Gemini.<\/p>\n\n\n\n

Ki\u1ebfn tr\u00fac Transformer v\u00e0 c\u01a1 ch\u1ebf Self-Attention<\/strong><\/h3>\n\n\n\n

Transformer ch\u00ednh l\u00e0 n\u1ec1n t\u1ea3ng c\u1ee7a m\u1ecdi m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef l\u1edbn hi\u1ec7n nay. C\u01a1 ch\u1ebf Self-Attention cho ph\u00e9p m\u00f4 h\u00ecnh x\u00e1c \u0111\u1ecbnh m\u1ee9c \u0111\u1ed9 li\u00ean quan gi\u1eefa c\u00e1c t\u1eeb trong m\u1ed9t c\u00e2u, t\u1eeb \u0111\u00f3 hi\u1ec3u ng\u1eef c\u1ea3nh ch\u00ednh x\u00e1c h\u01a1n nhi\u1ec1u so v\u1edbi c\u00e1c ki\u1ebfn tr\u00fac m\u1ea1ng n\u01a1-ron c\u0169. Ng\u01b0\u1eddi h\u1ecdc n\u00ean d\u00e0nh th\u1eddi gian t\u00ecm hi\u1ec3u kh\u00e1i ni\u1ec7m Token h\u00f3a (Tokenizer), c\u1eeda s\u1ed5 ng\u1eef c\u1ea3nh (Context Window) v\u00e0 c\u00e1ch t\u00ednh chi ph\u00ed s\u1eed d\u1ee5ng theo Token, v\u00ec \u0111\u00e2y l\u00e0 nh\u1eefng y\u1ebfu t\u1ed1 \u1ea3nh h\u01b0\u1edfng tr\u1ef1c ti\u1ebfp \u0111\u1ebfn chi ph\u00ed v\u1eadn h\u00e0nh khi tri\u1ec3n khai \u1ee9ng d\u1ee5ng th\u1ef1c t\u1ebf.<\/p>\n\n\n

\n
\"Giai
Giai \u0111o\u1ea1n 1 \u2013 L\u00e0m quen v\u1edbi LLMs<\/strong><\/figcaption><\/figure>\n<\/div>\n\n\n

Prompt Engineering \u2013 k\u1ef9 n\u0103ng giao ti\u1ebfp hi\u1ec7u qu\u1ea3 v\u1edbi LLMs<\/strong><\/h3>\n\n\n\n

Prompt Engineering l\u00e0 k\u1ef9 thu\u1eadt thi\u1ebft k\u1ebf c\u00e2u l\u1ec7nh sao cho m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef tr\u1ea3 v\u1ec1 k\u1ebft qu\u1ea3 ch\u00ednh x\u00e1c, \u0111\u00fang \u0111\u1ecbnh d\u1ea1ng mong mu\u1ed1n. K\u1ef9 n\u0103ng n\u00e0y bao g\u1ed3m vi\u1ec7c \u0111\u01b0a v\u00ed d\u1ee5 minh h\u1ecda (Few-shot Prompting), y\u00eau c\u1ea7u m\u00f4 h\u00ecnh tr\u00ecnh b\u00e0y t\u1eebng b\u01b0\u1edbc suy lu\u1eadn (Chain-of-Thought), v\u00e0 ki\u1ec3m so\u00e1t \u0111\u1ecbnh d\u1ea1ng \u0111\u1ea7u ra b\u1eb1ng JSON ho\u1eb7c XML. Ng\u01b0\u1eddi m\u1edbi n\u00ean luy\u1ec7n t\u1eadp vi\u1ebft prompt h\u1eb1ng ng\u00e0y v\u1edbi nhi\u1ec1u t\u00ecnh hu\u1ed1ng kh\u00e1c nhau \u0111\u1ec3 h\u00ecnh th\u00e0nh tr\u1ef1c gi\u00e1c v\u1ec1 c\u00e1ch m\u00f4 h\u00ecnh ph\u1ea3n h\u1ed3i.<\/p>\n\n\n\n

Giai \u0111o\u1ea1n 2 \u2013 N\u1eafm v\u1eefng k\u1ef9 thu\u1eadt RAG v\u00e0 qu\u1ea3n l\u00fd b\u1ed9 nh\u1edb cho LLMs<\/strong><\/h2>\n\n\n\n

Sau khi hi\u1ec3u n\u1ec1n t\u1ea3ng LLMs, b\u01b0\u1edbc ti\u1ebfp theo trong l\u1ed9 tr\u00ecnh h\u1ecdc LLMs & AI Agent l\u00e0 h\u1ecdc c\u00e1ch k\u1ebft n\u1ed1i m\u00f4 h\u00ecnh v\u1edbi d\u1eef li\u1ec7u ri\u00eang, thay v\u00ec ch\u1ec9 d\u1ef1a v\u00e0o ki\u1ebfn th\u1ee9c \u0111\u00e3 \u0111\u01b0\u1ee3c hu\u1ea5n luy\u1ec7n s\u1eb5n.<\/p>\n\n\n\n

Embeddings v\u00e0 Vector Database<\/strong><\/h3>\n\n\n\n

Embeddings l\u00e0 k\u1ef9 thu\u1eadt chuy\u1ec3n v\u0103n b\u1ea3n th\u00e0nh c\u00e1c vector s\u1ed1 h\u1ecdc, cho ph\u00e9p m\u00e1y t\u00ednh so s\u00e1nh m\u1ee9c \u0111\u1ed9 t\u01b0\u01a1ng \u0111\u1ed3ng v\u1ec1 ng\u1eef ngh\u0129a gi\u1eefa c\u00e1c \u0111o\u1ea1n v\u0103n b\u1ea3n kh\u00e1c nhau. Nh\u1eefng vector n\u00e0y \u0111\u01b0\u1ee3c l\u01b0u tr\u1eef trong c\u00e1c c\u01a1 s\u1edf d\u1eef li\u1ec7u vector chuy\u00ean d\u1ee5ng nh\u01b0 Chroma, Qdrant, Milvus ho\u1eb7c pgvector. M\u1ed7i c\u00f4ng c\u1ee5 c\u00f3 th\u1ebf m\u1ea1nh ri\u00eang, t\u1eeb \u0111\u01a1n gi\u1ea3n d\u1ec5 b\u1eaft \u0111\u1ea7u \u0111\u1ebfn kh\u1ea3 n\u0103ng m\u1edf r\u1ed9ng \u1edf quy m\u00f4 l\u1edbn.<\/p>\n\n\n

\n
\"Giai
Giai \u0111o\u1ea1n 2 \u2013 N\u1eafm v\u1eefng k\u1ef9 thu\u1eadt RAG v\u00e0 qu\u1ea3n l\u00fd b\u1ed9 nh\u1edb cho LLMs<\/strong><\/figcaption><\/figure>\n<\/div>\n\n\n

X\u00e2y d\u1ef1ng h\u1ec7 th\u1ed1ng RAG ho\u00e0n ch\u1ec9nh<\/strong><\/h3>\n\n\n\n

RAG (Retrieval-Augmented Generation) l\u00e0 k\u1ef9 thu\u1eadt k\u1ebft h\u1ee3p gi\u1eefa t\u00ecm ki\u1ebfm th\u00f4ng tin v\u00e0 sinh v\u0103n b\u1ea3n, gi\u00fap m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef tr\u1ea3 l\u1eddi d\u1ef1a tr\u00ean t\u00e0i li\u1ec7u n\u1ed9i b\u1ed9 c\u1ee7a doanh nghi\u1ec7p thay v\u00ec “b\u1ecba” th\u00f4ng tin. Quy tr\u00ecnh c\u01a1 b\u1ea3n g\u1ed3m: chia nh\u1ecf t\u00e0i li\u1ec7u, chuy\u1ec3n th\u00e0nh embeddings, l\u01b0u v\u00e0o vector database, sau \u0111\u00f3 truy xu\u1ea5t \u0111o\u1ea1n v\u0103n b\u1ea3n li\u00ean quan nh\u1ea5t m\u1ed7i khi c\u00f3 c\u00e2u h\u1ecfi \u0111\u1ec3 \u0111\u01b0a v\u00e0o prompt cho m\u00f4 h\u00ecnh x\u1eed l\u00fd. \u0110\u00e2y l\u00e0 k\u1ef9 thu\u1eadt g\u1ea7n nh\u01b0 b\u1eaft bu\u1ed9c khi x\u00e2y d\u1ef1ng chatbot ch\u0103m s\u00f3c kh\u00e1ch h\u00e0ng ho\u1eb7c tr\u1ee3 l\u00fd tra c\u1ee9u n\u1ed9i b\u1ed9.<\/p>\n\n\n\n

Giai \u0111o\u1ea1n 3 \u2013 B\u01b0\u1edbc v\u00e0o th\u1ebf gi\u1edbi AI Agent: T\u1eeb l\u00fd thuy\u1ebft \u0111\u1ebfn th\u1ef1c h\u00e0nh<\/strong><\/h2>\n\n\n\n

Khi \u0111\u00e3 n\u1eafm ch\u1eafc LLMs v\u00e0 RAG, ng\u01b0\u1eddi h\u1ecdc ch\u00ednh th\u1ee9c b\u01b0\u1edbc sang ph\u1ea7n th\u00fa v\u1ecb nh\u1ea5t c\u1ee7a l\u1ed9 tr\u00ecnh h\u1ecdc LLMs & AI Agent: x\u00e2y d\u1ef1ng nh\u1eefng h\u1ec7 th\u1ed1ng c\u00f3 kh\u1ea3 n\u0103ng h\u00e0nh \u0111\u1ed9ng t\u1ef1 ch\u1ee7.<\/p>\n\n\n\n

AI Agent kh\u00e1c g\u00ec so v\u1edbi chatbot truy\u1ec1n th\u1ed1ng<\/strong><\/h3>\n\n\n\n

C\u00f3 th\u1ec3 h\u00ecnh dung s\u1ef1 kh\u00e1c bi\u1ec7t qua m\u1ed9t v\u00ed d\u1ee5 \u0111\u01a1n gi\u1ea3n: chatbot gi\u1ed1ng nh\u01b0 h\u1ecfi \u0111\u01b0\u1eddng m\u1ed9t ng\u01b0\u1eddi – b\u1ea1n h\u1ecfi c\u00e1ch \u0111i \u0111\u1ebfn qu\u00e1n c\u00e0 ph\u00ea, ng\u01b0\u1eddi \u0111\u00f3 ch\u1ec9 \u0111\u01b0\u1eddng r\u1ed3i b\u1ea1n ph\u1ea3i t\u1ef1 di chuy\u1ec3n. AI Agent l\u1ea1i gi\u1ed1ng nh\u01b0 c\u00f3 t\u00e0i x\u1ebf ri\u00eang, ch\u1ec9 c\u1ea7n y\u00eau c\u1ea7u \u0111\u01b0a \u0111\u1ebfn qu\u00e1n c\u00e0 ph\u00ea, h\u1ec7 th\u1ed1ng s\u1ebd t\u1ef1 m\u1edf b\u1ea3n \u0111\u1ed3, t\u00ednh to\u00e1n l\u1ed9 tr\u00ecnh, tr\u00e1nh k\u1eb9t xe v\u00e0 \u0111\u01b0a b\u1ea1n \u0111\u1ebfn t\u1eadn n\u01a1i. S\u1ef1 kh\u00e1c bi\u1ec7t n\u00e0y n\u1eb1m \u1edf kh\u1ea3 n\u0103ng t\u1ef1 l\u00ean k\u1ebf ho\u1ea1ch v\u00e0 th\u1ef1c thi nhi\u1ec1u b\u01b0\u1edbc li\u00ean ti\u1ebfp m\u00e0 kh\u00f4ng c\u1ea7n con ng\u01b0\u1eddi h\u01b0\u1edbng d\u1eabn t\u1eebng thao t\u00e1c.<\/p>\n\n\n

\n
\"Giai
Giai \u0111o\u1ea1n 3 \u2013 B\u01b0\u1edbc v\u00e0o th\u1ebf gi\u1edbi AI Agent: T\u1eeb l\u00fd thuy\u1ebft \u0111\u1ebfn th\u1ef1c h\u00e0nh<\/strong><\/figcaption><\/figure>\n<\/div>\n\n\n

\ud83d\udc49 Xem th\u00eam: AI Agent l\u00e0 g\u00ec?<\/a><\/p>\n\n\n\n

V\u00f2ng l\u1eb7p Agent v\u00e0 c\u00e1c m\u00f4 h\u00ecnh ki\u1ebfn tr\u00fac ph\u1ed5 bi\u1ebfn<\/strong><\/h3>\n\n\n\n

Ph\u1ea7n l\u00f5i c\u1ee7a m\u1ecdi AI Agent l\u00e0 v\u00f2ng l\u1eb7p x\u1eed l\u00fd g\u1ed3m ba b\u01b0\u1edbc: quan s\u00e1t (nh\u1eadn th\u00f4ng tin t\u1eeb m\u00f4i tr\u01b0\u1eddng ho\u1eb7c ng\u01b0\u1eddi d\u00f9ng), suy lu\u1eadn (l\u00ean k\u1ebf ho\u1ea1ch h\u00e0nh \u0111\u1ed9ng ti\u1ebfp theo) v\u00e0 h\u00e0nh \u0111\u1ed9ng (g\u1ecdi c\u00f4ng c\u1ee5, th\u1ef1c thi t\u00e1c v\u1ee5). M\u1ed9t s\u1ed1 m\u00f4 h\u00ecnh ki\u1ebfn tr\u00fac ph\u1ed5 bi\u1ebfn ng\u01b0\u1eddi h\u1ecdc c\u1ea7n n\u1eafm g\u1ed3m ReAct (k\u1ebft h\u1ee3p suy lu\u1eadn v\u00e0 h\u00e0nh \u0111\u1ed9ng xen k\u1ebd), Planner-Executor (t\u00e1ch ri\u00eang vai tr\u00f2 l\u1eadp k\u1ebf ho\u1ea1ch v\u00e0 th\u1ef1c thi), c\u00f9ng c\u00e1c m\u00f4 h\u00ecnh ph\u1ed1i h\u1ee3p \u0111a Agent nh\u01b0 tu\u1ea7n t\u1ef1, song song, ph\u00e2n c\u1ea5p v\u00e0 chuy\u1ec3n giao nhi\u1ec7m v\u1ee5 gi\u1eefa c\u00e1c t\u00e1c nh\u00e2n.<\/p>\n\n\n\n

Giai \u0111o\u1ea1n 4 \u2013 L\u00e0m ch\u1ee7 c\u00e1c framework x\u00e2y d\u1ef1ng AI Agent ph\u1ed5 bi\u1ebfn <\/strong><\/h2>\n\n\n\n

Sau khi hi\u1ec3u l\u00fd thuy\u1ebft, ng\u01b0\u1eddi h\u1ecdc c\u1ea7n ch\u1ecdn c\u00f4ng c\u1ee5 ph\u00f9 h\u1ee3p \u0111\u1ec3 hi\u1ec7n th\u1ef1c h\u00f3a \u00fd t\u01b0\u1edfng. \u0110\u00e2y l\u00e0 b\u01b0\u1edbc quan tr\u1ecdng trong l\u1ed9 tr\u00ecnh h\u1ecdc LLMs & AI Agent, quy\u1ebft \u0111\u1ecbnh t\u1ed1c \u0111\u1ed9 tri\u1ec3n khai s\u1ea3n ph\u1ea9m th\u1ef1c t\u1ebf.<\/p>\n\n\n\n

LangGraph \u2013 ki\u1ec3m so\u00e1t lu\u1ed3ng x\u1eed l\u00fd ph\u1ee9c t\u1ea1p<\/strong><\/h3>\n\n\n\n

LangGraph l\u00e0 framework m\u00f4 h\u00ecnh h\u00f3a quy tr\u00ecnh Agent d\u01b0\u1edbi d\u1ea1ng \u0111\u1ed3 th\u1ecb tr\u1ea1ng th\u00e1i, ph\u00f9 h\u1ee3p v\u1edbi nh\u1eefng workflow c\u00f3 nh\u00e1nh r\u1ebd, v\u00f2ng l\u1eb7p th\u1eed l\u1ea1i ho\u1eb7c b\u01b0\u1edbc ph\u00ea duy\u1ec7t c\u1ee7a con ng\u01b0\u1eddi (Human-in-the-loop). \u0110\u00e2y l\u00e0 l\u1ef1a ch\u1ecdn \u0111\u01b0\u1ee3c nhi\u1ec1u t\u1ed5 ch\u1ee9c l\u1edbn tin d\u00f9ng nh\u1edd kh\u1ea3 n\u0103ng ki\u1ec3m so\u00e1t chi ti\u1ebft v\u00e0 truy v\u1ebft lu\u1ed3ng x\u1eed l\u00fd, tuy \u0111\u01b0\u1eddng cong h\u1ecdc t\u1eadp c\u00f3 ph\u1ea7n d\u1ed1c h\u01a1n so v\u1edbi c\u00e1c framework kh\u00e1c.<\/p>\n\n\n\n

\ud83d\udc49 Xem th\u00eam: \u0110\u00e1nh gi\u00e1 chi ti\u1ebft top 10+ AI Agent t\u1ed1t nh\u1ea5t<\/a><\/p>\n\n\n

\n
\"Giai
Giai \u0111o\u1ea1n 4 \u2013 L\u00e0m ch\u1ee7 c\u00e1c framework x\u00e2y d\u1ef1ng AI Agent ph\u1ed5 bi\u1ebfn <\/strong><\/figcaption><\/figure>\n<\/div>\n\n\n

CrewAI \u2013 x\u00e2y d\u1ef1ng \u0111\u1ed9i nh\u00f3m Agent theo vai tr\u00f2<\/strong><\/h3>\n\n\n\n

CrewAI ti\u1ebfp c\u1eadn theo h\u01b0\u1edbng \u0111\u01a1n gi\u1ea3n h\u01a1n: ng\u01b0\u1eddi d\u00f9ng \u0111\u1ecbnh ngh\u0129a c\u00e1c Agent v\u1edbi vai tr\u00f2, m\u1ee5c ti\u00eau v\u00e0 b\u1ed1i c\u1ea3nh ri\u00eang, gi\u1ed1ng nh\u01b0 ph\u00e2n c\u00f4ng c\u00f4ng vi\u1ec7c cho t\u1eebng th\u00e0nh vi\u00ean trong m\u1ed9t \u0111\u1ed9i nh\u00f3m. V\u00ed d\u1ee5 th\u1ef1c t\u1ebf th\u01b0\u1eddng g\u1eb7p l\u00e0 t\u1ea1o m\u1ed9t nh\u00f3m g\u1ed3m Agent nghi\u00ean c\u1ee9u th\u00f4ng tin, Agent vi\u1ebft n\u1ed9i dung v\u00e0 Agent bi\u00ean t\u1eadp – ph\u1ed1i h\u1ee3p t\u1ef1 \u0111\u1ed9ng \u0111\u1ec3 t\u1ea1o ra m\u1ed9t b\u00e0i vi\u1ebft ho\u00e0n ch\u1ec9nh. Nh\u1edd \u0111\u1ed9 kh\u00f3 th\u1ea5p, CrewAI th\u01b0\u1eddng l\u00e0 l\u1ef1a ch\u1ecdn \u0111\u1ea7u ti\u00ean cho ng\u01b0\u1eddi m\u1edbi mu\u1ed1n d\u1ef1ng th\u1eed nghi\u1ec7m nhanh.<\/p>\n\n\n\n

Claude Agent SDK v\u00e0 c\u00e1c l\u1ef1a ch\u1ecdn kh\u00e1c<\/strong><\/h3>\n\n\n\n

Ngo\u00e0i hai framework tr\u00ean, h\u1ec7 sinh th\u00e1i c\u00f4ng c\u1ee5 x\u00e2y d\u1ef1ng Agent hi\u1ec7n nay c\u00f2n c\u00f3 Claude Agent SDK (framework ch\u00ednh th\u1ee9c c\u1ee7a Anthropic, c\u00f9ng ki\u1ebfn tr\u00fac \u0111ang v\u1eadn h\u00e0nh Claude Code, m\u1ea1nh v\u1ec1 t\u00edch h\u1ee3p c\u00f4ng c\u1ee5 v\u00e0 giao th\u1ee9c MCP), AutoGen c\u1ee7a Microsoft (thi\u00ean v\u1ec1 l\u1eadp tr\u00ecnh h\u1ed9i tho\u1ea1i gi\u1eefa nhi\u1ec1u Agent), c\u00f9ng c\u00e1c l\u1ef1a ch\u1ecdn kh\u00e1c nh\u01b0 Semantic Kernel, LlamaIndex Workflows v\u00e0 Pydantic AI. L\u1ef1a ch\u1ecdn framework n\u00ean d\u1ef1a tr\u00ean nhu c\u1ea7u th\u1ef1c t\u1ebf: c\u1ea7n ki\u1ec3m so\u00e1t ch\u1eb7t ch\u1ebd ch\u1ecdn LangGraph, c\u1ea7n t\u1ed1c \u0111\u1ed9 tri\u1ec3n khai nhanh ch\u1ecdn CrewAI, c\u1ea7n t\u00edch h\u1ee3p s\u00e2u v\u00e0o h\u1ec7 sinh th\u00e1i Anthropic ch\u1ecdn Claude Agent SDK.<\/p>\n\n\n\n

Giai \u0111o\u1ea1n 5 \u2013 Tri\u1ec3n khai AI Agent v\u00e0o c\u00f4ng vi\u1ec7c v\u00e0 doanh nghi\u1ec7p th\u1ef1c t\u1ebf<\/strong><\/h2>\n\n\n\n

Giai \u0111o\u1ea1n cu\u1ed1i c\u00f9ng trong l\u1ed9 tr\u00ecnh h\u1ecdc LLMs & AI Agent t\u1eadp trung v\u00e0o vi\u1ec7c \u0111\u01b0a s\u1ea3n ph\u1ea9m t\u1eeb m\u00f4i tr\u01b0\u1eddng th\u1eed nghi\u1ec7m sang v\u1eadn h\u00e0nh th\u1ef1c t\u1ebf.<\/p>\n\n\n\n

K\u1ebft h\u1ee3p Function Calling v\u00e0 giao th\u1ee9c MCP<\/strong><\/h3>\n\n\n\n

Function Calling l\u00e0 c\u01a1 ch\u1ebf cho ph\u00e9p m\u00f4 h\u00ecnh ng\u00f4n ng\u1eef g\u1ecdi c\u00e1c h\u00e0m ho\u1eb7c API b\u00ean ngo\u00e0i \u0111\u1ec3 l\u1ea5y d\u1eef li\u1ec7u th\u1eddi gian th\u1ef1c, ch\u1eb3ng h\u1ea1n tra c\u1ee9u th\u1eddi ti\u1ebft, truy v\u1ea5n c\u01a1 s\u1edf d\u1eef li\u1ec7u ho\u1eb7c g\u1eedi email. <\/p>\n\n\n\n

MCP (Model Context Protocol) l\u00e0 giao th\u1ee9c m\u1edf \u0111ang tr\u1edf th\u00e0nh ti\u00eau chu\u1ea9n k\u1ebft n\u1ed1i gi\u1eefa Agent v\u1edbi c\u00f4ng c\u1ee5 v\u00e0 ngu\u1ed3n d\u1eef li\u1ec7u b\u00ean ngo\u00e0i, gi\u00fap c\u00e1c Agent x\u00e2y d\u1ef1ng t\u1eeb framework kh\u00e1c nhau v\u1eabn c\u00f3 th\u1ec3 giao ti\u1ebfp v\u00e0 ph\u1ed1i h\u1ee3p v\u1edbi nhau m\u1ed9t c\u00e1ch nh\u1ea5t qu\u00e1n.<\/p>\n\n\n

\n
\"Giai
Giai \u0111o\u1ea1n 5 \u2013 Tri\u1ec3n khai AI Agent v\u00e0o c\u00f4ng vi\u1ec7c v\u00e0 doanh nghi\u1ec7p th\u1ef1c t\u1ebf<\/strong><\/figcaption><\/figure>\n<\/div>\n\n\n

Gi\u00e1m s\u00e1t, \u0111\u00e1nh gi\u00e1 v\u00e0 t\u1ed1i \u01b0u hi\u1ec7u su\u1ea5t Agent<\/strong><\/h3>\n\n\n\n

Khi \u0111\u01b0a Agent v\u00e0o v\u1eadn h\u00e0nh th\u1ef1c t\u1ebf, vi\u1ec7c gi\u00e1m s\u00e1t l\u00e0 b\u1eaft bu\u1ed9c \u0111\u1ec3 \u0111\u1ea3m b\u1ea3o an to\u00e0n v\u00e0 hi\u1ec7u qu\u1ea3. Ng\u01b0\u1eddi h\u1ecdc n\u00ean l\u00e0m quen v\u1edbi c\u00e1c c\u00f4ng c\u1ee5 quan s\u00e1t h\u1ec7 th\u1ed1ng (Observability) nh\u01b0 Langfuse ho\u1eb7c LangSmith \u0111\u1ec3 theo d\u00f5i t\u1eebng b\u01b0\u1edbc x\u1eed l\u00fd, c\u00f9ng c\u00e1c c\u00f4ng c\u1ee5 \u0111\u00e1nh gi\u00e1 ch\u1ea5t l\u01b0\u1ee3ng \u0111\u1ea7u ra nh\u01b0 DeepEval ho\u1eb7c Promptfoo. M\u1ed9t l\u01b0u \u00fd quan tr\u1ecdng:<\/strong> giai \u0111o\u1ea1n th\u1eed nghi\u1ec7m n\u00ean c\u00f3 s\u1ef1 gi\u00e1m s\u00e1t ch\u1eb7t ch\u1ebd c\u1ee7a con ng\u01b0\u1eddi \u1edf b\u01b0\u1edbc \u0111\u1ea7u, sau \u0111\u00f3 t\u0103ng d\u1ea7n m\u1ee9c \u0111\u1ed9 t\u1ef1 \u0111\u1ed9ng h\u00f3a khi \u0111\u1ed9 tin c\u1eady \u0111\u01b0\u1ee3c ki\u1ec3m ch\u1ee9ng qua th\u1ef1c t\u1ebf v\u1eadn h\u00e0nh.<\/p>\n\n\n\n

B\u1ea1n c\u00f3 th\u1ec3 b\u1eaft \u0111\u1ea7u b\u1eb1ng c\u00e1c d\u1ef1 \u00e1n sau:<\/strong><\/p>\n\n\n\n