AI agent, đọc lại cho đúngAI agents, read correctly
Ảnh gốc nói: từ phản hồi đơn giản đến hành động tự chủ, như một cái thang 5 bậc.The infographic says: from simple responses to autonomous action, as a 5-step ladder.
Không phải cái thang. Câu hỏi là: ai chọn bước tiếp theo?Not a ladder. The question is: who picks the next step?
Code của bạn, hay model? Năm khái niệm trong ảnh là năm cách trả lời câu hỏi đó, mỗi cách một giá. Cấp cao hơn không tốt hơn, chỉ đắt hơn và khó đoán hơn.Your code, or the model? The five concepts in the infographic are five answers to that question, each with a price. A higher level isn't better, only costlier and less predictable.
Năm cấp xếp chồng, cấp sau kế thừa cấp trước.Memory và tool độc lập nhau. Multi-agent là kiến trúc, không phải bậc năng lực.Five stacked levels, each inheriting the last.Memory and tools are independent. Multi-agent is an architecture, not a capability tier.
Cấp 1 là một agent.Một lần gọi LLM không vòng lặp, không tool thì chưa phải agent. Nó là nền móng.Level 1 is an agent.One LLM call with no loop and no tools isn't an agent. It's the foundation.
Cấp 5 học từ kết quả và tự cải thiện.Model không tự huấn luyện lại khi chạy. "Học" là ghi bài học vào memory rồi đọc lại.Level 5 learns from outcomes and improves itself.The model doesn't retrain at run time. "Learning" means writing lessons to memory and reading them back.
Agent tự chạy trọn chiến dịch marketing.Chưa đáng tin khi không giám sát. Agent mạnh nhất ở việc có tín hiệu kiểm chứng tự động, như chạy test.An agent runs a whole marketing campaign alone.Not reliable unsupervised. Agents are strongest where there's an automatic check, like running tests.
Chọn cái nàoWhich one to pick
Trước hết: có khi bạn không cần thứ nàoFirst: you may need none of them
Không cần LLM khi quy tắc viết ra được và kết quả phải giống hệt mỗi lần: tính phí ship, kiểm tra định dạng, chuyển trạng thái đơn. Một câu if rẻ hơn hàng nghìn lần và không bao giờ bịa.No LLM when the rules can be written down and output must be identical every time: shipping fees, format checks, order status. An if is thousands of times cheaper and never hallucinates.
Không cần agent khi các bước đã biết trước. Viết luồng bằng code, chỉ gọi LLM ở bước cần hiểu ngôn ngữ.No agent when the steps are known. Write the flow in code; call the LLM only where language is needed.
Thử trước: một lần gọi LLM, prompt rõ, 3–5 ví dụ, output JSON schema, 30–50 case kiểm thử. Chỉ thêm độ phức tạp khi bộ kiểm thử chỉ ra lỗi mà prompt không sửa được.Try first: one LLM call, a clear prompt, 3–5 examples, JSON-schema output, 30–50 test cases. Add complexity only when tests show failures prompting can't fix.
Quy tắc viết ra được, không cần hiểu ngôn ngữ?
├─ Có ──► Code thường. Không dùng LLM.
└─ Không
Một lần gọi LLM với prompt tốt đã đủ?
├─ Đủ ──► Gọi LLM đơn. Dừng ở đây.
└─ Chưa — thiếu gì?
├─ Thiếu thông tin người dùng / lịch sử / tài liệu riêng
│ └─► thêm MEMORY
├─ Cần dữ liệu sống hoặc tác động ra hệ thống khác
│ └─► thêm TOOL
│ Các bước có biết trước không?
│ ├─ Biết ──► Workflow: code cầm lái
│ └─ Không ─► AGENT TỰ CHỦ + ngân sách cứng
│ + duyệt hành động không đảo ngược
└─ Tách được thành nhánh độc lập, chạy song song,
và đáng giá gấp ~15× token?
├─ Có ──► MULTI-AGENT
└─ Không ─► giữ một agent
Can the rules be written down, no language understanding needed?
├─ Yes ──► Plain code. No LLM.
└─ No
Is one LLM call with a good prompt enough?
├─ Yes ──► Single LLM call. Stop here.
└─ No — what's missing?
├─ User info / history / private documents
│ └─► add MEMORY
├─ Live data, or acting on another system
│ └─► add TOOLS
│ Are the steps known in advance?
│ ├─ Yes ──► Workflow: code drives
│ └─ No ───► AUTONOMOUS AGENT + hard budgets
│ + approval for irreversible actions
└─ Splits into independent parallel branches,
and worth ~15× the tokens?
├─ Yes ──► MULTI-AGENT
└─ No ───► keep one agent
Một lần gọi: prompt vào, văn bản ra. Không trạng thái, không tool, không vòng lặp.One call: prompt in, text out. No state, no tools, no loop.
Đổi khả năng tự tìm thông tin và hành động lấy chi phí thấp, tốc độ cao và kết quả dự đoán được.Trades finding information and acting on its own for low cost, speed and predictable results.
Dùng choUse for
Phân loại ticket, tóm tắt tài liệu dán vào, trích xuất JSON từ email, dịch, FAQ nhỏ.Ticket routing, summarising pasted docs, email-to-JSON extraction, translation, small FAQs.
ĐượcGain
Rẻ nhất, nhanh nhất, dễ kiểm thử nhất, không tác dụng phụ.Cheapest, fastest, easiest to test, no side effects.
MấtLose
Không biết gì ngoài prompt; bịa khi thiếu thông tin; không hành động; không nhớ người dùng.Knows nothing beyond the prompt; invents when unsure; can't act; no memory of users.
Dấu hiệu:Tell:đầu vào đã chứa đủ mọi thứ cần để trả lời.the input already holds everything needed to answer.
L2
Memory
Năng lực bổ sungAdded capabilityTrong ảnh: Context-Aware AgentsInfographic: Context-Aware Agents
Hệ thống lấy đúng thông tin từ kho lưu trữ và nhét vào prompt. Model không nhớ; hệ thống nhắc lại cho nó.The system fetches the right facts from a store and puts them in the prompt. The model doesn't remember; the system reminds it.
Đổi sự đơn giản không trạng thái lấy câu trả lời theo bối cảnh riêng của từng người.Trades stateless simplicity for answers grounded in each person's own context.
Dùng choUse for
CSKH nhớ đơn hàng, hỏi đáp tài liệu nội bộ (RAG), trợ lý nhớ văn phong, gia sư nhớ lỗi hay gặp.Support that remembers orders, internal-docs Q&A (RAG), assistants that keep your voice, tutors that track mistakes.
ĐượcGain
Cá nhân hoá không cần fine-tune; cập nhật tài liệu là có hiệu lực ngay; trích được nguồn.Personalisation without fine-tuning; doc updates apply instantly; citable sources.
MấtLose
Truy xuất sai thì trả lời sai rất tự tin; context phình to; memory lỗi thời; rủi ro quyền riêng tư.Bad retrieval gives confident wrong answers; bloated context; stale memory; privacy risk.
Dấu hiệu:Tell:câu trả lời đúng phụ thuộc vào thông tin không có trong câu hỏi. Dữ liệu có cấu trúc thì dùng SQL, không cần vector DB.the right answer depends on facts not in the question. Structured data wants SQL, not a vector DB.
L3
Tool use
Năng lực bổ sungAdded capabilityTrong ảnh: Action AgentsInfographic: Action Agents
Model chọn tool và tham số; code của bạn thực thi rồi trả kết quả lại. Model không tự gọi API.The model picks a tool and arguments; your code runs it and returns the result. The model never calls APIs itself.
Đổi sự an toàn của một hệ chỉ biết nói lấy khả năng thay đổi thế giới thật.Trades the safety of a system that only talks for the power to change the real world.
Dùng choUse for
Đặt lịch họp, tra đơn và hoàn tiền dưới ngưỡng, cập nhật CRM, giá và tồn kho thời gian thực, chạy SQL.Scheduling, order lookups and small refunds, CRM updates, live prices and stock, running SQL.
ĐượcGain
Dữ liệu mới nhất; làm được việc thật; tính toán chính xác bằng code; có log từng lời gọi.Fresh data; real work done; exact maths in code; every call logged.
MấtLose
Tác dụng phụ không đảo ngược; prompt injection; chọn sai tool khi mô tả mơ hồ; chậm và tốn hơn.Irreversible side effects; prompt injection; wrong tool from vague descriptions; slower and costlier.
Dấu hiệu:Tell:yêu cầu cần dữ liệu đang thay đổi, hoặc kết thúc bằng một động từ thay đổi hệ thống khác.the request needs changing data, or ends in a verb that changes another system.
L4
Multi-agent
Điều phối vòng lặpLoop orchestrationTrong ảnh: Multi-Agent SystemsInfographic: Multi-Agent Systems
Nhiều agent, mỗi cái một context riêng, chạy song song dưới một orchestrator. Giá trị thật là song song và tổng context lớn, không phải "nhiều vai".Several agents, each with its own context, running in parallel under an orchestrator. The real value is parallelism and more total context, not "many roles".
Đổi chi phí và tính nhất quán lấy độ phủ rộng và tốc độ song song.Trades cost and consistency for breadth and parallel speed.
Dùng choUse for
Nghiên cứu nhiều hướng độc lập, review code theo tiêu chí tách biệt, rà khối tài liệu lớn hơn một context.Research along independent lines, code review by separate criteria, sweeping more documents than one context holds.
ĐượcGain
Nhanh hơn trên việc rộng; context sạch; Anthropic báo hơn agent đơn 90,2% trên eval nghiên cứu nội bộ.Faster on broad work; clean contexts; Anthropic reports +90.2% over a single agent on an internal research eval.
MấtLose
~15× token; sub-agent không thấy nhau nên mâu thuẫn; gỡ lỗi rất khó; kém ở việc phụ thuộc chặt.~15× tokens; sub-agents can't see each other and conflict; very hard to debug; weak on tightly coupled work.
Dấu hiệu:Tell:việc chia được thành nhánh không cần biết nhánh khác làm gì. Chia theo nhiệm vụ, đừng chia theo vai người.the work splits into branches that don't need each other. Split by task, not job title.
L5
Agent tự chủAutonomous agent
Điều phối vòng lặpLoop orchestrationTrong ảnh: Autonomous AgentsInfographic: Autonomous Agents
Nhận mục tiêu, tự lập kế hoạch, hành động, quan sát, lặp tới khi xong. Model cầm vòng lặp. "Tự cải thiện" nghĩa là ghi bài học vào memory, không phải tự huấn luyện lại.Takes a goal, plans, acts, observes, repeats until done. The model holds the loop. "Self-improving" means writing lessons to memory, not retraining.
Đổi khả năng kiểm soát và dự đoán lấy năng lực xử lý việc không vẽ được quy trình trước.Trades control and predictability for handling work whose process can't be drawn in advance.
Dùng choUse for
Coding agent sửa bug tới khi test pass, điều tra sự cố qua log, nghiên cứu mở, vận hành có kiểm tra tự động.Coding agents fixing bugs until tests pass, log-driven incident triage, open research, ops with automatic checks.
ĐượcGain
Xử lý số bước không biết trước; tự sửa từ phản hồi môi trường; người chỉ giữ điểm quyết định.Handles unknown step counts; self-corrects from feedback; people stay only at decision points.
MấtLose
Lỗi cộng dồn (95%/bước × 20 bước ≈ 36%); chi phí khó đoán; có thể báo xong khi chưa xong.Compounding errors (95%/step × 20 steps ≈ 36%); unpredictable cost; may claim done when it isn't.
Dấu hiệu:Tell:không vẽ được quy trình trước, nhưng tự động kiểm tra được kết quả đúng hay sai.you can't draw the process, but you can automatically check whether the result is right.
Bảng đánh đổiTrade-off table
Gọi đơnSingle call
Memory
Tool use
Multi-agent
Tự chủAutonomous
Ai chọn bướcWho picks the step
Code
Code
Code hoặc modelCode or model
Model
Model
Token tương đốiRelative tokens
~1×
1× + context
1–4×
~15×
~4×, không trần nếu không đặt~4×, unbounded without a cap
Độ trễLatency
Một lần gọiOne call
+ truy xuất+ retrieval
+ mỗi vòng tool+ each tool round
Song song, vẫn chậm hơn workflowParallel, still slower than a workflow
Dài, khó đoánLong, unpredictable
Dễ kiểm thửTestability
CaoHigh
KháFair
Trung bìnhMedium
ThấpLow
ThấpLow
Rủi ro tác dụng phụSide-effect risk
KhôngNone
Lộ dữ liệuData exposure
CaoHigh
Cao × số agentHigh × agents
Cao nhất × số bướcHighest × steps
Độ khó gỡ lỗiDebugging
ThấpLow
Trung bìnhMedium
Trung bìnhMedium
Rất caoVery high
CaoHigh
~4× và ~15× so với một lượt chat, theo Anthropic (06/2025). Dùng như thứ tự độ lớn, không phải hằng số.~4× and ~15× relative to a chat turn, per Anthropic (June 2025). Orders of magnitude, not constants.
Anti-patternsAnti-patterns
Sai lầmMistake
Vì sao saiWhy it's wrong
Sửa thế nàoFix
Coi 5 cấp là lộ trình phải leoTreating the 5 levels as a path to climb
Cấp cao hơn chỉ đắt hơn, không tốt hơnHigher is only costlier, not better
Chọn từng năng lực theo nhu cầuPick each capability by need
Agent cho quy trình cố địnhAn agent for a fixed process
Trả giá agent cho việc workflow làm rẻ và chắc hơnPaying agent prices for workflow work
Luồng bằng code, LLM ở bước cần ngôn ngữFlow in code, LLM where language is needed
Multi-agent từ ngày đầuMulti-agent on day one
~15× token, lỗi phối hợp khó truy~15× tokens, coordination bugs
Một agent trước, tách khi có số đoOne agent first, split on evidence
Chia agent theo vai ngườiSplitting agents by job title
Ranh giới vai ≠ ranh giới contextRole boundaries ≠ context boundaries
Chia theo nhánh việc độc lậpSplit by independent branches
Quyền ghi rộngBroad write access
Một lần hiểu sai = 1.000 email gửi nhầmOne misread = 1,000 wrong emails
Quyền tối thiểu, duyệt hành động không đảo ngượcLeast privilege, approve irreversible actions
Nhét toàn bộ lịch sử vào contextWhole history in context
Tốn tiền, model bỏ sót phần giữaCostly, middle gets missed
Truy xuất chọn lọc, tóm tắt định kỳSelective retrieval, periodic summaries
Tin agent "tự học"Believing agents "self-learn"
Không cập nhật trọng số, lặp lại lỗi cũNo weight updates, old mistakes repeat
Ghi bài học vào memory, có kiểm thửWrite lessons to memory, with tests
Coi output tool là mệnh lệnhTool output treated as commands
Prompt injection
Coi là dữ liệu, tách quyền đọc/ghiTreat as data, split read/write
Đánh giá bằng cảm giácJudging by feel
Không biết thay đổi tốt lên hay tệ điCan't tell better from worse
Bộ kiểm thử cố định + đọc traceFixed test set + trace review
Ba câu cần nhớThree things to remember
Agent không phải cái thang: memory, tool, nhiều agent và quyền tự chủ là những núm vặn độc lập, mỗi núm có giá riêng.Agents aren't a ladder: memory, tools, multiple agents and autonomy are independent dials, each with its own price.
Nếu vẽ được quy trình trước, hãy để code cầm vòng lặp.If you can draw the process in advance, let code hold the loop.
Bắt đầu từ một lần gọi LLM có bộ kiểm thử; chỉ thêm độ phức tạp khi số đo chứng minh nó đáng tiền.Start from one LLM call with a test set; add complexity only when measurements prove it pays.