AI Agent Development and Interview Guide: 00-Recruitment Requirements and Competency Model Report

1. Core Conclusions

AI Agent R&D has long moved beyond the stage of "only knowing how to invoke model APIs and write Prompts." The focus of the current recruitment market is how to connect large models with databases, internal tools, and business processes to build stable, observable, and measurable production-grade systems.

From this sample:

• RAG and retrieval capabilities have become foundational skills. All 21 samples mention RAG, knowledge bases, Embeddings, vector databases, or related retrieval technologies. Since this search centered on Agent positions, this ratio should not be extrapolated as a whole-market proportion, but it sufficiently proves that "chat-only Agents" are clearly insufficient.
• Production-grade software engineering and project evidence are almost hard thresholds. 20/21 mention APIs, concurrency, databases, caching, microservices, containers, cloud, CI/CD, monitoring, or production runtime; 21/21 require or prioritize verifiable experiences such as demonstrable projects, deployed systems, open-source contributions, or prototype demos.
• Tool calling and orchestration are the core of Agent positions. 18/21 mention Function Calling, Tool Use, MCP, or toolchains; 14/21 mention Multi-Agent, A2A, or complex workflows. Enterprises want "systems that can take action," not merely generate text.
• Evaluation and observability have entered mainstream requirements. 14/21 explicitly mention Eval, Tracing, logging, success rate, latency, cost, replay, or A/B testing. Senior positions especially emphasize golden datasets, automated evaluation, online/offline metrics, and regression testing.
• Python remains the preferred language, but not the only language. 16/21 explicitly name Python; Java, Go, TypeScript/Node.js often serve as enterprise backend, tool services, or full-stack capability supplements.
• Model fine-tuning is not the first priority for all positions. 12/21 mention Fine-tuning, LoRA, RLHF/DPO, inference deployment, or training pipelines, many of which are bonus qualifications. For most application-oriented positions, reliable orchestration, RAG, evaluation, and engineering are more critical than training models independently.
• Educational thresholds have obvious regional differences. 14/14 Chinese samples specify Bachelor's degree or above, and Baidu's Agent Algorithm campus recruitment requires Master's or above; overseas samples more often emphasize production experience and deliverables, and only Manus and Tessera explicitly state Bachelor's or above degree requirements among the 7 samples.
• Fresh graduates have opportunities, but must prove abilities with projects. 8/21 explicitly accept campus recruitment, excellent fresh graduates, ~1 year experience, or count personal projects/internships into experience; "at least one end-to-end project that can be demonstrated live and explain technical trade-offs" is the most realistic admission ticket.

2. Research Methodology and Boundaries

2.1 Sample Selection

Included samples are all positions whose job titles contain AI Agent, Intelligent Agent, Agent R&D, or whose job responsibilities explicitly center on building Agent systems:

• Covering campus recruitment, junior, 1–3 years, 3–5 years, as well as senior/expert positions;
• Covering Internet, e-commerce, finance, education, hardware manufacturing, enterprise services, cybersecurity, and startup teams;
• Sources include official company websites, recruitment ATS systems, campus hiring brochures, university employment portals, and public pages of mainstream hiring platforms.

2.2 Statistical Criteria

Frequencies only count explicit mentions in job descriptions (including "Job Responsibilities" and "Qualifications/Bonus Points"), and do not represent a linear proportion of absolute importance in actual work.

2.3 Limitations

• Belongs to targeted purposive sampling, not randomized market-wide sampling, and cannot be used to deduce industry-wide percentages;
• Recruitment information is time-sensitive and may be adjusted or closed at any time; this capture date is July 22, 2026 (2026-07-22);
• Platform displayed information is subject to final employer communication; the Huaqin position comes from a third-party aggregation page and serves only as supplementary reference;
• Different positions have widely varying depth requirements for the same term: "familiar with LangChain" may just mean basic usage, or it may require reading source code and secondary development.

3. Requirement Frequencies

Interpretation Focus: Lower frequency does not mean unimportant. Security governance accounts for only 48% in all samples, but is more concentrated in finance, enterprise software, and senior overseas positions, often serving as a critical threshold after systems enter production environments.

4. Capability Structure Truly Needed by Employers

4.1 Tier 1: Solid Software Engineering Foundations

Typical requirements include Python, data structures and algorithms, asynchronous programming, API design, databases, caching, message queues, microservices, and distributed systems. PDD requires high concurrency, high availability, caching, message queues, and asynchronous scheduling; Xiaocheng Group further specifies CAP, consistency, Kafka/RabbitMQ, CI/CD, Kubernetes, and canary releases; Syntronic requires scalable, highly available systems, testing, CI/CD, performance, and cost control.

This indicates that Agent engineering does not replace traditional backend capabilities, but rather adds model non-determinism, tool side effects, and evaluation challenges on top of them.

4.2 Tier 2: Agent Core Mechanisms

High-frequency modules include:

• Planning, routing, reflection, state management, and short-term/long-term memory;
• Function Calling, Tool Use, MCP, plugins, and Skills;
• ReAct, Plan-and-Execute, Multi-Agent collaboration, A2A;
• Failure retries, checkpoints, recovery, degradation, and Human-in-the-Loop;
• Context injection, compression, caching, long text, and Token budgets.

Positions such as PDD, Baidu, Syntronic, Planera, etc. treat planning, tools, memory, evaluation, or security as system modules, rather than regarding framework names themselves as goals.

4.3 Tier 3: RAG and Knowledge Engineering

Enterprise expectations have upgraded from "connecting a vector database" to a full retrieval pipeline:

• Document parsing and cleaning;
• Chunking strategies and metadata;
• Embedding selection;
• Hybrid retrieval with vectors, BM25, or knowledge graphs;
• Rerank, Query Rewrite, and context filtering;
• Citations, traceability, and hallucination control;
• RAGAS, DeepEval, or custom dataset evaluation.

The requirements of UnionPay Gold Card, Kelibang, and Xiaocheng Group all cover this complete pipeline. If candidates only write "familiar with RAG" without recall, ranking, evaluation, and failure cases, it is difficult to create differentiation.

4.4 Tier 4: Evaluation, Observability, and Reliability

Agent outputs possess non-determinism, so employers are starting to demand "designing evaluation before continuous deployment." Common metrics include:

• Task success rate, tool calling success rate, and answer correctness;
• Retrieval recall, faithfulness, and citation accuracy;
• Latency, Token consumption, invocation cost, and business conversion;
• Tracing, structured logging, error classification, replay, and regression testing;
• A/B comparisons between Prompts, tools, and model versions.

Planera explicitly requires golden datasets, automated evaluation, snapshot replay, and online/offline quality metrics; Lumos emphasizes tracing, alerting, and the entire Agent lifecycle; PDD incorporates success rate, latency, Token, and business conversion rate into its observability system.

4.5 Tier 5: Security and Side-Effect Governance

Production Agents will read and write databases, invoke external systems, and even execute code. Senior roles therefore demand:

• Prompt Injection and malicious input protection;
• Tool whitelisting, parameter validation, least privilege, authorization, and auditing;
• Code sandboxes and high-risk action confirmation;
• PII/sensitive data handling, content moderation, and compliance;
• Timeout, retry, idempotency, compensation, and rollback strategies.

Security is not a separate "hardening phase," but a design constraint for tool interfaces and execution workflows.

4.6 Tier 6: Business Processes and Product Outcomes

20/21 samples require business understanding, cross-team collaboration, or end-to-end outcomes. Syntronic requires breaking complex processes into executable atomic tasks; educational positions require abstracting pedagogical processes jointly with teachers, curriculum researchers, and product teams; overseas positions such as TRM, Arcade, Planera, etc. emphasize user impact, rapid experimentation, and end-to-end ownership.

Therefore, excellent candidates need to explain not just "what framework was used," but also: why use an Agent, why not use a deterministic workflow, how to degrade upon failure, by how much metrics improved, and whether the system genuinely saved business time.

5. Four Common Role Profiles

Job titles cannot fully represent job content. When job hunting, judge whether a position leans more toward algorithms, platforms, applications, or industry delivery based on the JD body, then adjust project and resume emphasis.

6. Degree and Experience Thresholds

6.1 Chinese Positions

• 14/14 explicitly require Bachelor's degree or above; Baidu's Agent Algorithm campus hiring requires Master's or above.
• Campus hiring positions still demand LLM/Agent principles, RAG, frameworks, and project practice, and are not zero-foundation trainee roles.
• 1–3 years positions typically require independently completing prototypes or 0 to 1 (0–1) projects.
• 3–5 years positions emphasize production experience, architecture, performance, cost, evaluation, security, and cross-team driving.

6.2 Overseas Positions

• Degree frequencies are lower, while production system experience, code quality, product impact, and ownership are more prominent.
• Startups often demand full-stack capability: Agent backend, tool layer, evaluation platform, and even React/TypeScript UI.
• Senior positions typically require 4–5+ years of software engineering experience, but not necessarily multiple years of "Agent seniority"; strong backend/distributed backgrounds are transferable.

6.3 Realistic Judgment for Fresh Graduates

Fresh graduates can enter this direction, but need to satisfy three points simultaneously:

1. Have solid programming and computer science foundations;
2. Have at least one complete, runnable Agent project that can explain technical trade-offs;
3. The project not only demonstrates the happy path, but also includes evaluation, exception handling, logging, and deployment.

7. Differences Between Chinese and Overseas Samples

The commonality between both markets is even more important: both are shifting from "model callers" to "reliable Agent system builders."

8. Portfolio Standards Corresponding to Recruitment Requirements

Rather than building multiple shallow Demos, it is better to build one publicly demonstrable production-grade project. Recommended works include the following:

8.1 Must Demonstrate

• At least 3 real tools, using structured input and output Schemas;
• Stateful workflow, including routing, retries, timeouts, degradation, and human confirmation;
• RAG using hybrid search and Rerank, returning verifiable citations;
• FastAPI or equivalent backend, including async calls, database/caching, and error handling;
• Docker one-click startup, with README clarifying architecture, boundaries, and run methods;
• A fixed set of evaluation data, outputting task success rate, quality, latency, and cost;
• Tracing/logging capable of pinpointing whether failures occur in planning, retrieval, tool, or generation stages.

8.2 Differentiating Content

• MCP Server or reusable Skill/Tool SDK;
• Prompt Injection testing, permission models, and sandbox execution;
• Model routing, caching, parallel invocations, and Token cost optimization;
• Automated replay, offline evaluation, and CI regression gates;
• Real user feedback, business time saved, or conversion metrics;
• Written post-mortems of failure cases, rather than only displaying successful screenshots.

8.3 Resume Expression Template

Do not just write "built Agent using LangChain." A more effective phrasing is:

Designed and launched a LangGraph-based multi-step business Agent, integrating 5 permission-controlled tools and hybrid search RAG; established a 120-case golden evaluation dataset with automated replay, improving task success rate from 68% to 87%, reducing P95 latency by 31%, and lowering per-task Token cost by 24%.

Numbers must come from real measurements; when online data is unavailable, report results from fixed benchmark test sets while noting the experimental setup.

9. Interview Preparation Checklist

Inferred from job requirements, technical interviews have a high probability of revolving around the following questions:

1. Why does this scenario need an Agent, rather than a normal workflow or single LLM call?
2. How do you design Tool Schema, permissions, idempotency, retries, and side-effect confirmation?
3. How do you choose Chunk size, Embedding, Hybrid Search, and Rerank?
4. How do you construct evaluation datasets to distinguish between retrieval failures, reasoning failures, and tool failures?
5. How do you reduce latency, Tokens, and model costs, while avoiding significant quality degradation?
6. When does Multi-Agent provide ROI, and when does it merely increase complexity?
7. How do you prevent Prompt Injection, unauthorized tool calls, and unsafe code execution?
8. How do you make the system support replay, tracing, version comparison, and production regression?
9. What was the most severe failure case in the project, and how was it ultimately isolated and fixed?
10. How do you decompose ambiguous business goals into quantifiable Agent success criteria?

10. Recommended 8-Week Preparation Roadmap

11. Sample Positions and Sources

Chinese Positions (14)

1. PDD Holdings | AI Agent R&D Engineer: Core architecture, RAG, high concurrency, security, observability, MCP, Multi-Agent.
2. Tianjin Xiaocheng Group | AI Agent Development Engineer (AI Native): Python, distributed systems, RAG, Eval, containers, and end-to-end projects.
3. Alibaba Group | AI Agent Campus Recruitment: Python, Transformer, deep learning; Agent frameworks, RAG, and open source as bonus points.
4. Baidu | 2027AIDU Intelligent Agent Algorithm Engineer: Planning, tool calling, Multi-Agent, RAG, evaluation, Fine-tuning/RL.
5. Syntronic | LLM Agent Engineer: Business process decomposition, Skills, Context/Harness Engineering, observability, and security.
6. Kaidu Investment | Large Model Application Development Engineer: Coding Agent, RAG, MCP, Skills, Docker, and business process transformation.
7. Guangxi Guisheng Financial IT | AI R&D Engineer: Financial AI project full lifecycle, Function Calling, MCP, RAG, Dify/Coze/n8n.
8. Beijing UnionPay Gold Card Tech | System Development Engineer (LLM Application Direction): Full RAG chain, Agent components, RAGAS, FastAPI, Docker.
9. Nanjing BaiAo | AI Agent Senior Development Engineer: Python, FastAPI, LangGraph, Multi-Agent, Graph RAG, Docker.
10. Kelibang | Agent Development Engineer: Agent architecture, RAG, Prompt, cost/latency, robustness, private deployment.
11. Shanghai Linglingyi | Junior AI Agent Development Engineer: Coze/Dify, LangChain/LangGraph, RAG, A2A, and AI Coding.
12. Huaqin Technology | AI Agent Development Engineer: Task planning, tool calling, Multi-Agent, inference optimization, and local deployment; this link is a third-party aggregation page.
13. Senzong Aishu | AI Intelligent Agent Application Senior Development Engineer: Educational Agent, Coze/HiAgent, platform integration, data evaluation, security compliance.
14. Zhengzhou Yuehao | Large Model Application Development Engineer Campus Recruitment: Python/Java, RAG, Prompt, MCP, cloud product support.

Overseas Positions (7)

15. Manus AI | AI Agent Engineer: Agent core modules, Prompt, Tool, RAG, evaluation, and rapid prototyping.
16. Lumos | AI Agent Engineer: Agent SDK, safety fallback, Tracing, alerting, governance, and full-stack engineering.
17. TRM Labs | AI Agent Engineer: Python backend, LLM/Agent, vector databases, security, scalability, and product empathy.
18. Tessera Labs | AI Engineer: Enterprise Multi-Agent, tool calling, model training/evaluation, and cloud infrastructure.
19. Arcade | Agent Engineer: Prompt/Tool Chain, Trace, A/B, offline evaluation, product experience, and full-stack collaboration.
20. Atria | AI Agent Engineer: Multi-Agent, memory, AgentOps, Agentic RAG, Docker/Kubernetes, and security governance.
21. Planera | Senior AI Agent Engineer: LangGraph, MCP, golden dataset, replay, observability, sandbox, and multi-model routing.

12. Final Judgment

If all recruitment requirements are compressed into a single sentence, what the market needs is:

A person who can use solid software engineering to orchestrate LLMs, retrieval, and tools into reliable Agents, and prove with evaluation data that it is effective in real business.

For job seekers, the priorities should be:

1. First build up Python/backend and systems engineering foundations;
2. Then master RAG, Tool Use/MCP, and stateful Agent orchestration;
3. Integrate Eval, Tracing, security, cost, and deployment into projects;
4. Finally supplement fine-tuning, reinforcement learning, multi-modal, or vertical industry capabilities based on target positions.

Frameworks will change rapidly, but "problem decomposition — reliable execution — quantitative evaluation — business outcomes" is the more enduring core competency.