The Agent Handbook

Appearance

JD List: Live Roles at Home and Abroad

At a glance About 20 live Agent-track JDs compiled from real searches in August 2026: domestic coverage includes ByteDance, Alibaba, Tencent, Meituan, and MiniMax; overseas includes OpenAI, Anthropic, Cursor, Cognition, and Sierra—each with a source link, ending with a word-frequency summary of high-frequency JD requirements.

This page contains time-sensitive content; data is current as of 2026-08. Job listings, pricing, and product features may have changed — verify against the original sources before citing.

JD List: Live Roles at Home and Abroad ​

This page does no theoretical analysis and does exactly one thing: put the Agent-track jobs actually posted, at home and abroad, as of August 2026 in front of you. Every entry carries a source link you can click through to read the original—the JD itself always contains more information than any secondhand summary.

1. Read This First: How to Use This List ​

Three usage principles:

  1. A JD is a capability benchmark, not a wish list. Don't count how many requirements you meet; find the one that "appears across multiple JDs and you can't do"—that's your next learning priority. Works best paired with the Knowledge Map.
  2. Mind the timestamps. The hiring market changes weekly. Entries marked "live" were confirmed posted at search time (August 2026); entries marked "historical JD, for capability benchmarking" may have their original links taken down, but the capability descriptions remain valid. Always re-confirm on the company's official careers page before applying.
  3. Read the requirements, not the title. "AI Agent Engineer," "LLM Application Engineer," and "Agent Infra Engineer" may be the same job at different companies; conversely, two roles both titled "Agent Engineer"—one building products, one building RL environments—are entirely different species. The judging criterion is always the keywords in the JD body.

One judgment

The 2026 JD market has clearly stratified: engineering roles building Agent products/harnesses (high volume, comparatively lower bar), research roles doing Agent training/RL (low volume, top-conference papers or strong engineering background required), and FDE roles delivering Agents to customers (surging both domestically and abroad). First decide which layer you're entering, then study against those JDs—don't blast one resume everywhere.

2. Domestic Roles ​

ByteDance ​

ByteDance currently has the densest concentration of Agent roles in China, hiring across ads, data platforms, and security. The following come from BOSS Zhipin searches (showing "live" or recently updated as of August 2026):

1. AI Agent Engineer – Big Data Development Suite (Beijing/Shanghai)

  • Team: Data Platform Development Suite team, building intelligence into the unified big-data R&D governance platform
  • Keywords: Golang / Java / C++ / Python, Agent, RAG, LLM
  • Core requirement: deep understanding of the data production pipeline, using LLMs to reshape the data engineering and governance experience; a typical "business + Agent" hybrid role
  • Source: BOSS Zhipin job page

2. Agentic RL R&D Engineer – Ads Infra (Beijing)

  • Keywords: reinforcement learning, Agentic, ads infra, bachelor's and up, any experience level
  • Note: a training-side Agent role, belonging to the "research/training layer" described above
  • Source: BOSS Zhipin job page

3. AI Agent Security Engineer Expert – Intelligent Security (Beijing/Hangzhou/Shenzhen, 35-65K × 15, 5-10 years)

  • Keywords: full-chain AI security, Agent security, red-teaming/defense
  • Note: Agent security is turning from a "bonus item" into a standalone role—nearly invisible in 2024 JDs
  • Source: BOSS Zhipin ByteDance AI job list

Alibaba / Alibaba Cloud ​

Alibaba's Agent roles concentrate on two lines—Alibaba Cloud and DingTalk—with a strong To-B flavor and a general 16-month pay structure:

4. Alibaba Cloud – AI Agent R&D Engineer (Hangzhou, 30-60K × 16, 3-5 years)

5. ATH Business Group – AI Agent R&D Expert – AI Video Production (Beijing, 50-80K × 16)

  • Keywords: Agent + multimodal/video production pipeline
  • Note: the salary band marks this as an expert role; "Agent + vertical content production" is a common big-tech combination in 2026
  • Source: BOSS Zhipin job list

6. Alibaba Cloud Intelligence – AI Agent Test Development Engineer/Expert (Hangzhou, 40-70K × 16, 3-5 years)

  • Keywords: full-lifecycle quality assurance for Agent products, test system construction
  • Note: Agent evaluation and QA have become standalone roles, directly matching what Evaluation Systems covers—an underrated entry direction
  • Source: BOSS Zhipin job list

Tencent ​

Tencent's Agent roles span WXG (WeChat), WeCom, Hunyuan, Yuanbao, and other teams; the official channel is careers.tencent.com:

7. WeChat – LLM (Agent) Application Backend Engineer (Beijing, 2+ years)

8. WeCom – Agent Algorithm Engineer (Guangzhou, 3+ years)

  • Core responsibilities: building multi-agent systems with autonomous decision-making, collaboration, and tool-calling capabilities; applying complex reasoning techniques such as CoT and ToT; instruction fine-tuning, reinforcement learning
  • Keywords: Multi-Agent, CoT/ToT, SFT, RL
  • Note: a rare domestic algorithm JD that spells out multi-agent and reasoning techniques in concrete detail
  • Source: Niuqi Zhipin repost (source: Tencent official site)

Meituan ​

9. Senior Java Development Engineer (AI Agent track) (Shanghai, 5-10 years)

10. Senior Agent Development Engineer (Full-Stack) (Beijing, 3-5 years)

  • Keywords: full-stack, Agent product development
  • Note: local-life scenarios (in-store, food delivery, customer service) are natural soil for Agent adoption
  • Source: BOSS Zhipin job page

MiniMax / Moonshot AI / Zhipu ​

11. MiniMax – Agent Test Development Engineer (2026 campus hiring)

  • Core responsibilities: AI product test design, defect tracking, test framework development and maintenance, building CI pipelines
  • Requirements: new bachelor's/master's graduates, CS-related majors, proficiency in at least one programming language (e.g. Python)
  • Note: friendly to new graduates—a low-friction entry into a top model company
  • Source: Nowcoder campus-hiring job page (application window 2025-12 to 2027-01; still open at search time)

12. Moonshot AI – Kimi Agent-track Internship (historical JD, for capability benchmarking)

  • Note: the Kimi Agent internship appeared on hiring channels previously; the official careers page (moonshot.cn/careers) was accessible at search time, but the job list loads dynamically, so current openings couldn't be verified item by item
  • Source: Moonshot AI official site (defer to live listings)

13. Zhipu (Z.ai) – LLM/Agent-related roles (historical JD, for capability benchmarking)

  • Note: Zhipu's "Join Us" page was accessible at search time, but specific listings couldn't be verified item by item. Per third-party tallies, Zhipu's hiring skews toward B2B delivery (government, energy, and financial clients), with Agent delivery roles tied to industry solutions
  • Source: Zhipu official "Join Us"

3. Overseas Roles ​

Overseas roles defer to official careers pages. Every entry below was verified against the official site in August 2026.

OpenAI ​

14. Software Engineer, Codex Core Agents (San Francisco + 3 other locations)

  • Team: Codex – Engineering, OpenAI's coding Agent product line
  • Note: at search time, OpenAI's careers site listed 27 live roles matching the keyword "agent" (out of 736 total live roles), spanning Agent Infrastructure, Agent Post-Training (with multiple tracks including Computer Use, Context, Connectors, Frontier Evals and Environments), Security Engineer, Agent Security, Software Engineer, API Agents, and more
  • Observation: the bulk of OpenAI's Agent roles sit in post-training and security—once the product form stabilized, firepower concentrated on "making Agents more reliable"
  • Source: OpenAI Careers (agent search results)

Anthropic ​

15. Staff+ Software Engineer, Claude Managed Agents (San Francisco/New York)

  • Note: Claude Managed Agents is Anthropic's managed Agent product line; this is a core engineering role building the product directly
  • Source: Anthropic Careers (live at search time)

16. Tech Lead Manager, Agent Runtime Platform (San Francisco/New York)

  • Note: the Agent runtime platform—a TLM role running the Agent execution infrastructure, corresponding to the domestic "Agent Infra" direction
  • Source: same as above, Anthropic Careers

17. Staff Software Engineer, Claude Code / Research Engineer, Computer Use (multiple locations)

  • Note: the Claude Code team and the Computer Use research team are hiring simultaneously; Anthropic's site also listed a large number of Forward Deployed Engineer (FDE) roles covering SF/NY/London/Paris/Munich
  • Source: same as above, Anthropic Careers

Cursor (Anysphere) ​

18. Software Engineer, Agent Harness (San Francisco/New York)

  • Core responsibilities: building the core harness that supports Agent behavior—the execution skeleton of the Agent (exactly what the Agent Loop chapter covers)
  • Source: Cursor Careers (live at search time)

19. Software Engineer, Agent Evaluation and Quality (San Francisco/New York)

  • Core responsibilities: building the Agent's measurement and quality systems
  • Note: Cursor is also hiring an Engineering Manager, Agent & Product Security. A coding-Agent company standing up "evaluation quality" and "Agent security" as separate engineering teams says these are now standard equipment for Agent products, not nice-to-haves
  • Source: same as above, Cursor Careers

Sierra ​

20. Agent Engineer / Agent Engineer, TLM (New York, TLM salary band $280K–$410K)

  • Core responsibilities: owning the full Agent Development Life Cycle (ADLC), designing, building, and delivering production-grade Agents directly to enterprise customers
  • Keywords (from the JD): eval frameworks, agent tooling, RAG pipelines, prompt engineering, production systems, React / TypeScript / Go
  • Note: Sierra is the origin of the "Agent Engineer" title (see its official blog in the references); for anyone trying to understand the FDE model, this JD is the best specimen
  • Source: Sierra Careers; salary band per ZipRecruiter aggregation page

Cognition (Devin) ​

21. Member of Technical Staff / General Application (San Francisco)

  • Note: Cognition's careers page was accessible at search time; per third-party tallies (July 2026) it had 78+ full-time openings, with SWE roles centered on Devin, the autonomous coding Agent. Apply via the site's General Application
  • Keywords: production coding, unfamiliar-codebase navigation, safety and reliability of autonomous agents
  • Source: Cognition Careers

4. Word-Frequency Summary: The 15 Most Common JD Requirements ​

A manual tally based on the JD bodies collected on this page (including full-text reposts found in searches), grouped into three tiers by how widely each requirement appears. This is a qualitative ranking, not an exact count—but the tier structure is very stable, so you can safely use it to order your learning priorities.

TierRequirement keywordsNotes
Tier 1 (in nearly every JD)1. LLM/LLM application experienceNot "used the API"—a complete project shipped to production
2. PythonThe de facto standard on both sides of the Pacific
3. Prompt EngineeringStill a baseline skill, but no longer a standalone job
4. RAGSee the RAG chapter; mandatory for delivery roles
5. Tool calling / Function CallingThe interface layer between Agent and systems; see Tools & MCP
Tier 2 (in most JDs)6. Agent architecture / task planning / multi-turn decision-makingi.e. the ability to design the Agent Loop
7. Production-grade systems experience (HA/distributed)The shared foundation of domestic backend roles and overseas SWE roles
8. Go / Java / TypeScript (at least one)Domestic skews Go/Java; overseas skews TypeScript/Go
9. Evaluation / Eval / Quality assuranceStandalone roles at Cursor, Alibaba Cloud, MiniMax
10. Fine-tuning / Post-Training / RLThe ticket into training-layer roles
Tier 3 (high-frequency pluses)11. Multi-Agent systemsWritten explicitly into JDs like Tencent WeCom
12. MCP and other tool/communication protocolsMoving from "aware of" to "familiar with"
13. CI/CD and testing frameworksThe core of Agent engineering-quality roles
14. Client communication / delivery skillsA hard requirement for FDE-type roles
15. Agent security / red-teamingA new standalone direction in 2026

Three counterintuitive conclusions

  1. "Prompt Engineer" as a standalone job is dead, yet prompt engineering as a skill appears in nearly every JD—it has retreated from a profession back into a skill.
  2. The experience bar is lower than you think. Epoch AI's tally of 1,604 Chinese AI job postings found Chinese companies require only 1.6 years of experience on average (US frontier labs: 5.5)—domestic Agent roles weigh work samples and fundamentals far more than years served; letting your portfolio projects speak beats grinding seniority.
  3. Evals are the most underrated entry point. For people who want into Agents but can't out-compete algorithm roles, the Agent evaluation/quality direction has visibly lower competition, and its skill requirements (designing evals, building pipelines) are fully self-teachable—see Evals in Practice.

5. Market Sketch: Three Signals from August 2026 ​

  • Domestic: three cities split the map. Epoch AI data shows Beijing, Hangzhou, and Shanghai absorbed 93% of location-tagged roles (Beijing alone 63%). Of ByteDance's 5,000+ campus openings for 2026, roughly 90% are AI-related.
  • Overseas: the FDE model blooms everywhere. Anthropic, Cursor, and Sierra are all hiring Forward Deployed Engineers at scale—the bottleneck of Agent companies has shifted from "building models" to "wedging Agents into customer businesses," mirroring the growth of domestic To-B delivery roles.
  • Research side: the harness becomes a discipline. DeepSeek is assembling a Harness team to rival Claude Code, and Cursor has a dedicated Agent Harness role—"scaffolding beyond the model" is now recognized as core competitive strength, exactly the point Anatomy of an Agent keeps making.

Next-step advice: pick the 3 JDs above you most want to reach, match the keywords against yourself one by one, find the corresponding exam points in the interview question list for your gaps, then rewrite your project descriptions against the resume analysis.

References ​