Sr. AI Agent Engineer 資深 AI Agent 工程師

Position Summary

我們正在組建 AI Agent 工程師的團隊，團隊目前的職責有三個方向：
1. 內部軟體工程流程自動化——建置能自動化工程流程的 Agent，包括日誌分析、問題分類與分派（triage）、自動化故障排除，以及給開發者用的工具。
2. 商家營運自動化——建置處理商家端營運流程的 Agent，能與使用者互動以蒐集缺漏資訊，並在案件超出其處理範圍時，完整交接給人工團隊。
3. 對客戶端的 AI 產品——作為產品本身一部分（而非內部工具）推出的 AI 功能：商家直接互動的 Agent、以自然語言存取營運資料、AI 輔助設定，以及產品內的協助功能。這個方向最需要嚴謹的設計：延遲與成本是需要被慎重評估的，且每一次 prompt 或工具變更都必須先經過版本控管與評估，才能上線給實際商家使用。

你會負責其中一到兩件事，不會全都做。負責哪些在到職後決定，或依你的專長和團隊當下的優先順序而定。

We are building out an AI Agent team, and this is not a “do everything” seat. The team’s charter currently spans three tracks:

• Internal SWE workflow automation — agents that automate engineering workflows: log anal-ysis, triage, automated troubleshooting, and developer-facing tooling.

• Merchant operations automation — agents that handle merchant-facing operational work-flows, interactively gather missing information from users, and escalate cleanly to human teams when a case exceeds their scope.

• Customer-facing AI product — AI capabilities that ship as part of the product itself rather than as internal tooling: agents merchants interact with directly, natural-language access to their own operational data, AI-assisted setup and configuration, and in-product assistance. This track carries the tightest constraints of the three. Latency and cost are visible to the user, there is no internal operator standing by to catch a bad response, and every prompt or tool change needs to be versioned and evaluated before it reaches live merchants.

You will own one to two of these tracks, not all three. Which ones you take is decided during onboarding, based on your strengths and the team’s priorities at the time. We are hiring additional engineers for the remaining scope.

Responsibilities

• 負責正式環境 Agent 的完整工程生命週期：多步驟規劃、工具設計、互動式資訊蒐集、確定性的備援流程（fallback），以及交接給人工團隊的路徑。
• 把目前靠人工處理的流程自動化，讓重複性的營運負擔不再持續耗費大量工程師的工時。
• 研究、評測並做出 Agent 模式的原型 — 自訂執行迴圈、反思與自我修正機制、動態重試策略、評估框架與沙盒環境 — 並依實際量測結果決定要不要採用，而不是因為它新。
• 建置評估與追蹤（tracing）層，讓 Agent 的可靠性可以被量化。
• 將 Agent 與既有的後端服務及內部 API 整合。
• 在跨職能的 Scrum 團隊工作，與橫跨印尼、台灣的 PM、前後端工程師與 SQA 協作。把零散的營運流程整理成結構化的 Agent 工具和工作流程。

Tech Stack

Strong Knowledge

• Agent 架構與編排——LangGraph，或自建的 Python Agent 引擎。我們不限定框架，重點在於明確的狀態管理、清楚的工具邊界，以及可測試性。包含工具/函式呼叫（tool/function calling）與多輪互動迴圈。
• LLM 與上下文管理——實際用過 SOTA 模型（Anthropic Claude、OpenAI、Gemini）、Model Context Protocol（MCP）、結構化輸出解析、動態提示（dynamic prompting）與 RAG 架構。向量資料庫（pgvector、Qdrant 或Pinecone）熟其中一種即可。
• Python 作為 AI/Agent 技術堆疊的主要程式語言。
• 評估與可觀測性——實作評估框架（DeepEval、LLM-as-a-judge；檢索相關評估則用 Ragas）以及 Agent 追蹤工具（LangSmith、Langfuse），以量化任務成功率、可靠性與幻覺（hallucination）發生率。
• LLM 系統的可靠性工程——深入了解各種失效模式（幻覺、推理迴圈、API 速率限制、schema 不一致），並實際設計過確定性備援、自我修正與交接機制。
• 多 Agent 模式——Supervisor/worker 架構、專職子 Agent，以及 Human-in-the-Loop 的執行控制。

Good Knowledge

• Java / Spring Boot——能看懂我們既有的服務、跟內部 API 整合即可。寫新的 Spring 服務不會是你的主要工作。
• 系統架構——RESTful/gRPC API 設計、用於長時間非同步 Agent 任務的 Redis 與訊息佇列（message queue）、關聯式資料庫（MySQL）。
• 安全與防護機制——AI 安全風險，包括提示注入（prompt injection）、敏感資料外洩、憑證隔離；外部 API 的速率限制策略。
• 雲端與 DevOps——Docker、Kubernetes（GKE），以及用於部署容器化 Agent 服務的 CI/CD 流程。
• 成本與延遲優化——上下文視窗裁剪、Prompt 快取（prompt caching），以及在前沿模型與較小型/自架模型（例如透過 vLLM 提供服務）之間進行任務路由。
  

Engineering Culture

• Production readiness——PoC 是起點，不是交付物。任何進到正式環境的東西，都必須安全、可觀測，而且原作者以外的人也看得懂且可維護。
• 在沙盒中快速迭代，在正式環境中求穩——隨時開實驗、用沒碰過的框架、砍掉重來都可以。但當你要把一個流程或工具推上正式環境時，就要通過評估和討論(RFC 或是內部討論)，依照狀態處理、失效模式與影響範圍（blast radius）等做審查，過關之前追求短回饋循環，過關之後追求可預測性。
• Clean code 與模組化的工具設計——高內聚、低耦合，工具應設計為可獨立測試的單元。

Responsibilities

• Own the full engineering lifecycle of production agents: multi-step planning, tool design, inter-active information gathering, deterministic fallback, and escalation paths to human teams.
• Automate workflows that currently depend on manual firefighting, so that recurring operational load stops consuming engineering time.
• Research, benchmark, and prototype emerging agentic patterns — custom execution loops, reflection and self-correction mechanisms, dynamic retry strategies, evaluation harnesses and sandboxes — and make adoption decisions based on measured results rather than novelty.
• Build the evaluation and tracing layer that makes agent reliability measurable rather than anec-dotal.
• Integrate agents with existing backend services and internal APIs.
• Work in a cross-functional scrum team alongside PMs, Backend/Frontend engineers, and SQA across Indonesia and Taiwan. Take fragmented operational processes and turn them into structured agent tools and clear technical specifications.

Tech Stack

Strong Knowledge

• Agent architecture & orchestration — LangGraph, or a custom Python agent engine. We are framework-agnostic; what matters is explicit state management, clean tool boundaries, and testability. Tool/function calling and multi-turn interactive loops.
• LLMs & context management — Hands-on with SOTA models (Anthropic Claude, OpenAI, Gemini), Model Context Protocol (MCP), structured output parsing, dynamic prompting, and RAG architectures. Depth in one vector store (pgvector, Qdrant, or Pinecone) is sufficient.
• Python as your primary language for the AI/agent stack.
• Evals & observability — Implementing evaluation frameworks (DeepEval, LLM-as-a-judge; Ra-gas for retrieval-specific evaluation) and agent tracing (LangSmith, Langfuse) to quantify task success, reliability, and hallucination rate.
• Reliability engineering for LLM systems — Deep understanding of failure modes (hallucination, reasoning loops, API rate limits, schema mismatch) and proven experience designing determin-istic fallback, self-correction, and escalation mechanisms.
• Multi-agent patterns — Supervisor/worker structures, specialized sub-agents, and Human-in-the-Loop execution controls.

Good Knowledge

• Java / Spring Boot — enough to read our existing services and integrate with internal APIs.
• Writing new Spring services will not be your main job.
• System architecture — RESTful/gRPC API design, Redis and message queues for long-running asynchronous agent tasks, relational databases (MySQL).
• Security & guardrails — AI security risks including prompt injection, sensitive data leakage, and credential isolation; rate-limiting strategies for external APIs.
• Cloud & DevOps — Docker, Kubernetes (GKE), and CI/CD pipelines for deploying containerized agent services.
• Cost & latency optimization — Context window trimming, prompt caching, and routing tasks between frontier models and smaller or self-hosted models (e.g. served via vLLM).

Engineering Culture

• Production readiness. A PoC is the starting point, not the deliverable. Anything that reaches production is expected to be secure, observable, and maintainable by someone other than its author.
• Fast in the sandbox, deliberate in production. Prototyping is intentionally unconstrained — spin up experiments, try unfamiliar frameworks, throw them away. The RFC gate applies when you propose a flow or tool for production, where design review covers state handling, failure modes, and blast radius. Short feedback loops before that gate; predictability after it.
• Clean code and modular tool design. High cohesion, low coupling, and tools designed as independently testable units.

Requirements

經驗

• 5 年以上專業軟體開發經驗，並具備扎實的後端系統工程、安全性與效能優化基礎。
• 具備 LLM 應用、AI Agent 或工作流程自動化工具的正式環境經驗，大約一年以上。這個領域仍相當年輕，我們更看重「你實際做出了什麼」而非「你做了多久」：比起一個花三年只做過純提示型聊天機器人的人，我們更想與一位花八個月時間、在真實使用場景下維持一個 Agent 穩定運作的人聊聊。
• 我們看重的是：你曾經營運過一個「真的有人依賴」的 LLM 系統，並且能具體描述它是如何出問題、你又做了哪些應對與改進。

工程品質與問題解決能力

• 系統性除錯——能夠除錯複雜的多步驟工具鏈、跨日誌追蹤問題，並診斷系統整合方面的問題。
• 能夠處理模糊性——能夠接手一個定義模糊的營運問題，評估其背後的流程，並提出一個範疇明確、可執行的 Agent 設計方案。

心態

• 工程務實精神與愛動手嘗試的直覺——重視系統的可預測性、資料隱私，以及成本/延遲效率，而非追逐 AI 熱潮。享受透過建置端到端 PoC 來驗證真實的工程可行性。
• 持續學習——主動關注 Agent 生態系的發展，並以「實務上是否值得採用」而非單純追新的角度來評估新工具。

溝通能力

• 具備良好的跨國協作能力（與印尼及台灣的同事合作）。
• 英文書寫與口說能力足以進行日常工作溝通與文件撰寫。

加分條件

• 具備獨立建置 LLM/Agent side project 或 PoC 的實績。
• 曾有 AI 輔助程式撰寫工具、自動化日誌分析器，或 auto-triage 系統的相關經驗。
• 曾有互動式問答機器人的相關經驗，該機器人能在將邊緣案例升級給客服或工程團隊之前，先向使用者蒐集缺漏資訊。
• 具備在正式環境中營運 self-hosted inference 的經驗。

Experience

• 5+ years of professional software development experience, with a solid foundation in backend system engineering, security, and performance optimization.
• Production experience with LLM-powered applications, AI agents, or workflow automation tools roughly a year or more. This field is young enough that we weigh what you shipped over how long you shipped it: we would rather talk to someone who spent eight months keeping one agent alive under real usage than someone who spent three years on prompt-only chatbots.
• What we are looking for is evidence you have operated an LLM system that people actually depended on, and that you can describe concretely how it broke and what you changed in response.

Engineering quality & problem solving

• Systemic debugging — Able to debug complex multi-step toolchains, trace across logs, and diagnose system integration issues.
• Comfortable with ambiguity — Able to take a vaguely defined operational problem, evaluate the underlying process, and come back with a scoped, executable agent design.

Mindset

• Engineering pragmatism and a tinkerer’s instinct — Cares about system predictability, data privacy, and cost/latency efficiency over AI hype. Enjoys building end-to-end PoCs to establish real engineering feasibility.
• Continuous learning — Actively follows the agent ecosystem and evaluates new tooling for practical adoption, not for its own sake.

Communication

• Strong cross-border collaboration skills (colleagues across Indonesia and Taiwan).
• Proficient written and spoken English for workplace communication and documentation.

Nice to Have

• A track record of independently built LLM/agent side projects or PoCs.
• Prior experience with AI-assisted coding tools, automated log analyzers, or auto-triage systems.
• Prior experience with interactive Q&A bots that collect missing information from users before escalating edge cases to support or engineering.
• Experience operating self-hosted inference in production.