跪拜 Guibai
← Back to the summary

DeepSeek's AI Agent Job Posting Is a Three-Month Engineering Map

Original Job Posting Link

https://app.mokahr.com/social-recruitment/high-flyer/140576#/job/8d40c764-d2b2-49b1-826c-e3f2adb75c01


image.png

This job requirement, on the surface, is a position description. In essence, it is a capability map. It implies the superposition of two layers of ability:

Many people easily confuse these two layers, resulting in learning that is both scattered and shallow, mastering neither.

Below, we proceed in four steps:

  1. First, deconstruct the original JD to see clearly what it is really testing.
  2. Then, explain the technical terms one by one, following the actual execution order of an Agent, clarifying the principles and the reasons behind the designs together.
  3. Next, list a complete technology stack panorama, grounding concepts into specific code tools by layer.
  4. Finally, provide a complete learning path with a timeline, implementation projects, and acceptance criteria.

You can treat it as a three-to-six-month self-study map, or as a checklist for filling gaps before an interview.


1. First, See Clearly What the JD is Really Asking For

Deconstructing the original text into two parts makes the logic clearer:

Part 1 (Engineering Capability)

Ability to proficiently use AI Agent tools to develop software, and to write quality-assured code with AI assistance even in a completely inexperienced technical domain.

This tests not "whether you can write code," but "whether you can use AI to quickly master an unfamiliar tech stack and guarantee output quality."

Part 2 (Principle Capability)

Requires you not only to be a heavy user of Agent products but also to understand the operating mechanisms behind these products—how LLMs are called, how context is managed, how tools are invoked, and how multiple Agents collaborate.

This tests "whether you know why these products are designed this way, and the reasons behind their usability."

Summary in One Sentence

The JD is not looking for "someone who can write code with AI," but rather "someone who can both deliver efficiently using Agents and clearly explain the internal principles of Agents, even capable of modifying the Agent's operating environment themselves."

There is a memory formula to help you贯穿 the entire text: Agent = LLM + Planning + Memory + Tools. All the terms that follow are essentially adding details to these four parts.


2. Explaining Technical Terms One by One (Principles and the 'Why')

The explanation below follows the actual execution order of an Agent system, not alphabetical order. This makes it easier to build a holistic picture rather than memorizing a pile of isolated concepts.

1. LLM API — The Starting Point of Everything

The LLM API is the interface through which you call a large model: you send a piece of text, and the model returns a piece of text.

Three key points beginners easily overlook:

All subsequent Agent mechanisms essentially solve the same problem: between successive LLM API calls, how to organize the content sent to the model and how to process the returned content. Once you thoroughly understand the API call, the following concepts become much easier.


2. KV Cache — The Hidden Mechanism Determining Agent Efficiency and Cost

This is the point in the JD most easily overlooked by beginners but highly valued by senior engineers.

Every time a large model generates a token, it needs the attention intermediate results (Key and Value, KV for short) calculated from all previous tokens. Recalculating from scratch every time is slow and costly. KV Cache caches these intermediate results. When the beginning of a new request is identical to the previous round, the cache is directly reused, and only the new part is calculated.

Practical impact: If you frequently insert or modify content at the beginning of the conversation history (like dynamically rewriting the system prompt), it causes frequent KV Cache invalidation, slowing down speed and increasing costs. This is why many Agent systems are designed with the principle of "only appending at the end, never modifying previously generated parts."

Understanding this point allows you to grasp why mainstream Agent products' context management strategies are designed that way—this is often a key question in interviews distinguishing between "having used" and "understanding the principles."


3. Agent Loop — The Heart of an Agent

An Agent is essentially a loop:

Receive task → Model thinks about the next step → Calls a tool to execute → Feeds the execution result back to the model → Model continues thinking → ... until the task is complete.

This pattern is often called ReAct (Reasoning + Acting).

The most common mistake beginners make is treating an Agent as "an advanced version that generates answers in one go." In reality, the key to an Agent is its ability to continuously adjust its strategy based on new information across multiple loops—this is the essential difference between an Agent and a regular chatbot. Without this loop, it's not an Agent.


4. Tool Use — The Hand with Which an Agent Interacts with the World

The model itself can only generate text; it cannot actually query a database, send an email, or execute code. Tool Use allows the model to "declare" its intention to call a specific tool and what parameters to pass when generating text, which an external program then actually executes, returning the result to the model.

The specific mechanism: when calling the API, you tell the model what tools are available (name, description, parameter format). When needed, the model returns a structured call request instead of plain text.

This is the technical foundation enabling all Agent products to "do things" rather than just "chat."


5. Reasoning + 6. Planning

These two concepts are closely related and are discussed together.

A point beginners easily overlook: Good Planning is not about generating a complete plan once and never changing it, but dynamically adjusting it based on execution results within the Agent Loop. This is why Planning and Agent Loop are tightly bound concepts.

Corresponding to the JD's requirement of "high-quality programming with AI even in inexperienced domains," it relies precisely on this planning and reasoning capability—decomposing the learning path for an unfamiliar tech stack and executing it step by step.


7. Skills — Packaging Experience into Reusable Knowledge Modules

Instead of stuffing all operational details into the system prompt, making the model "recall" them anew each time, it's better to write the best practices for a certain type of task into independent documents or script packages (like "how to generate a standard Word document"), allowing the Agent to load them on demand.

Benefits:


8. MCP (Model Context Protocol) — The Standard Interface for Agents to Connect to the External World

Before MCP emerged, connecting an Agent to Gmail, Slack, GitHub, or a database required writing separate adapter code for each service, and integrations couldn't be reused across different Agent products.

MCP is an open protocol defining a unified communication standard between Agents and external tools/data sources. It can be understood as "the USB interface of the Agent world" — as long as a service implements MCP, any MCP-supporting Agent client can connect directly, eliminating redundant wheel-reinvention.

This is one of the fastest-growing directions in the current Agent ecosystem and a necessary standard for building enterprise-grade Agent systems.


9. Memory — Allowing the Agent to Remember "What Happened Before"

The LLM itself does not remember any history; the required context must be passed back with each call. Memory mechanisms are typically divided into three layers:

Memory Type Content Stored Technical Carrier
Short-term/Working Memory Intermediate state of the current task In-Memory / Redis
Medium-term Memory Complete conversation history MySQL / PostgreSQL
Long-term Memory User preferences and factual knowledge retained across conversations Vector Database (Semantic Retrieval)

If you want to delve deeper into this layer, directly对照 the three-layer memory architecture you are familiar with is the best case study; no other教材 is needed.


10. Subagent + 11. Multi-Agent

In Stage 4 of the learning path, you will build such a Multi-Agent system using LangGraph, grounding the concept into engineering practice.


12. Prompt Engineering / Context Engineering / Harness Engineering — Three Progressive Layers, Boundaries Must Be Clear

These three concepts appear repeatedly in the JD and are the easiest to confuse. Their boundaries must be clearly distinguished:

Concept Scope of Concern Core Operations Plain Understanding
Prompt Engineering Input text for a single call Role setting, constraints, few-shot examples, output format control "Writing this one 'instruction' well"
Context Engineering Organization of all information during the entire run History trimming/compression, tool result summarization, RAG knowledge injection, Token budget allocation "Managing the multi-turn 'background information' well"
Harness Engineering The engineering architecture of the entire Agent system Loop design, tool registration/dispatching, error handling/retries, multi-Agent communication, state persistence, observability, human-in-the-loop intervention "Building the entire 'operating environment' well"

The JD specifically mentions "great passion for Agent Harness development," indicating the position truly values the third-layer capability—not being satisfied with using pre-built products, but having the interest to design and build this operating environment oneself.

The boundaries between these three layers are a high-frequency interview topic, easily追问 and easily confused. It is recommended to consciously think during each learning stage: Which of these three layers does what I am doing now belong to?


3. Complete Technology Stack Panorama

The previous section covered the "concept layer," but concepts must be grounded to land. This section lists the technology stack for an entire Agent system by layer, explaining what each layer is and why it is needed.

Layer 1: Programming Language & Backend Web Framework

Technology Role
Python 3.10+ Beyond core syntax, focus on supplementing Type Hints—not optional, as FastAPI and mainstream Agent frameworks heavily depend on them
FastAPI The mainstream choice for building Agent backend services, natively supports async (async/await), suitable for many I/O waiting scenarios
Uvicorn The ASGI server for running FastAPI; understand startup and --reload debugging

Layer 2: Frontend & Real-time Communication

Technology Role
HTML/CSS/JS Basics No need for mastery, but must be able to read DOM structure and basic JS syntax
Vue3 (Composition API) Component-based frontend UI development; learn the Composition API directly (setup, ref, reactive)
Vite The build tool配套 for Vue3; use npm create vue@latest to start a project
Axios / Native fetch The foundation for frontend calls to backend APIs
SSE (Server-Sent Events) Agent output streams token by token; frontend renders a "typewriter effect" in real-time. Simpler than WebSocket, it is the mainstream first choice

Layer 3: Agent Orchestration Frameworks

Technology Positioning
LangChain "Parts Library"—a pluggable, componentized tool library (Prompt templates, LLM calls, output parsers, tool wrappers)
LangGraph "Assembly Blueprint"—a state graph orchestration framework, natively supporting conditional branching, loops, and state persistence

Relationship between the two: LangChain provides the parts, LangGraph handles the assembly. They are not replacements for each other but are used together.

Layer 4: Data & Storage Layer

Technology Purpose
MySQL / PostgreSQL Store structured data like conversation history, user profiles (medium-term memory)
Redis Store intermediate states during Agent execution, short-term/working memory; fast read/write speed
Vector Database (Start with Chroma, advance to Qdrant/Milvus) Store long-term memory and RAG knowledge bases; retrieval based on semantic similarity

Layer 5: Communication & Protocols

Layer 6: Engineering & Deployment

Technology Purpose
Git Version management, essential for team collaboration
Docker Package the entire system (backend + database + vector database) into a portable container
Langfuse / LangSmith (Optional) Agent observability tools, tracking each step's call and execution time

4. A Learning Path for Beginners (Five Stages)

The following five stages progress according to the logic of "First Use → Then Understand → Finally Modify." Each stage includes: objectives, learning content, a落地 project, and which JD requirement it corresponds to.


Stage 0: Programming & Web Framework Basics

(Essential for absolute beginners, skippable if you have the basics, duration 1-2 weeks)

Objective: Acquire the minimum programming ability needed to read and call Agent systems, while getting early exposure to the web frameworks used later.

Learning Content:

Landing Project: A FastAPI-based endpoint that internally calls the LLM API to implement continuous multi-turn conversation, tested successfully with Postman or a simple script.


Stage 1: Become a Heavy User of AI Programming Tools + Supplement Frontend Basics

(Duration 1-2 weeks)

Objective: Make "using AI-assisted development" an instinctive reaction, meeting the JD's requirement of being a "high-intensity Agent product user." Simultaneously, use this opportunity to learn the unfamiliar tech stack of Vue3.

Specific Actions:

During the process, record how you asked AI questions and how you verified whether the Vue3 code given by AI was correct. First, establish a "trust but verify" work habit.

Landing Project: A small Vue3 + FastAPI frontend-backend integration project—the frontend is a simple dialog box that calls the backend FastAPI endpoint to display LLM Q&A results. Simultaneously, compile your own "AI-assisted development checklist."


Stage 2: Implement a Simplest Agent Loop by Hand

(Duration 2-3 weeks)

Objective: Stop treating the Agent as a black box and know what it actually does internally. Also, experience the difference between "handwriting" and "using a framework."

Specific Actions:

  1. Handwritten Version: Using only the raw LLM API, write the simplest Agent Loop—give the model a "calculator" tool and let it solve a problem requiring multiple tool calls.
  2. Add Planning: Have the model output a task decomposition list first, then execute item by item, observing how the code needs to handle plan changes mid-execution.
  3. Add Short-term Memory: Correctly pass the conversation history back in each round.
  4. Framework Refactoring: Rewrite the same logic using LangChain's Tool abstraction and Prompt templates, comparing the differences in code volume and structure between the two versions.

Landing Project: Two versions of the same Agent (pure handwritten version + LangChain version), supporting at least 2 tool calls, including short-term memory, capable of completing the full closed loop of "requirement decomposition → execution → verification," and write a comparison note explaining the pros and cons of both versions.


Stage 3: Three Major Engineering Practices + Introduce LangGraph and Storage Layer

(Duration 3-5 weeks, core assessment for the position)

This stage upgrades layer by layer in the order of Prompt → Context → Harness.

Step 1: Prompt Engineering (Approx. 1 week)

Write multiple versions of prompts for the same requirement for comparison testing, focusing on practicing controlling AI to write standard comments, write unit tests, and output in a specific format (JSON/Markdown code blocks). Produce a personal template library of code development prompts.

Step 2: Context Engineering (Approx. 1-2 weeks)

Add a layer of RAG to the Agent built in Stage 2:

Interlude: Principle Supplement (Parallel with the above two steps)

For each topic, write a 200-300 word explanation "to teach someone else"—only when you can write it out have you truly understood it.

Step 3: Harness Engineering (Approx. 2-3 weeks)

This is the focus of technology stack integration:

Landing Project: A complete frontend-backend integrated Agent system—Vue3 frontend + FastAPI backend + LangGraph orchestration + MySQL/Redis storage + Chroma vector retrieval, supporting streaming conversation, short-term and long-term memory, and capable of calling at least 2 tools.


Stage 4: Advanced System Development, Benchmarking the Position's Entry Standard

(Duration 2-4 weeks)

Objective: Evolve from "using Agent products" to "having the ability to design an Agent operating environment yourself."

Specific Actions:

  1. Multi-Agent System: Use LangGraph's multi-graph collaboration to build a Multi-Agent system, simulating an enterprise-level software development scenario—

    • Planning Agent decomposes requirements.
    • Backend Agent writes APIs.
    • Frontend Agent writes pages.
    • Testing Agent automatically writes unit tests.
    • Review Agent checks code quality.
    • Multiple Agents collaborate based on a clear protocol.
  2. Engineering Deployment: Package the entire system with Docker (backend service + MySQL + Redis + vector database, all containerized).

  3. Ultimate Verification: Choose a tech stack you have absolutely no experience with (like Go backend, React frontend, Mini Program development), and rely solely on this Multi-Agent system to autonomously plan, search, code, and debug to complete a full project, with manual intervention and debugging possible throughout.

This is the verification method directly benchmarking the JD requirement of "high-quality programming even in inexperienced domains," and it is the highest-value project for your resume.

Two things to do continuously:

If you have spare capacity, you can接入 Langfuse or similar observability tools to add execution trace tracking to the entire Multi-Agent system.


5. Learning Acceptance Checklist (Self-check if JD Requirements are Met)

After completing the above five stages, use this checklist for self-assessment:


6. A Reminder

What this JD essentially requires is not mastery of a specific framework's API, but a systematic understanding of "how LLMs are organized into a reliable system."

Technical terms and specific frameworks will change—today it's LangGraph, tomorrow it might be another framework; today it's Vue3, tomorrow it might be another frontend framework. But the underlying principles of Agent Loop, context management, and tool use are relatively stable.

It is recommended to focus your learning on "understanding principles + hands-on practice." The tech stack is merely the carrier for grounding principles, not the endpoint of learning. This way, even as tools iterate, your capabilities will not become obsolete.


7. A More Efficient Shortcut

The learning path above is a solid self-study map. But if you wish to not only traverse this path in a shorter time but also discuss with peers heading in the same direction and receive immediate feedback from senior engineers, you might consider another learning method.

The recently launched 「Harness & Hermes」 Multi-Agent Development Special Training Camp by Imooc has a curriculum outline whose design思路 highly aligns with the capability model mentioned in this article. It doesn't just teach API calls but centers on Harness Engineering as the core, guiding you to hand-write a complete Multi-Agent system from 0 to 1.

If interested, you can learn about Imooc's 「Harness & Hermes」 Multi-Agent Development Special Training Camp.