Appearance
Job Hunting & JD Analysis
What's the best compass for learning a technology? Not a tutorial's table of contents, not a list of papers — what the hiring market is actually paying for.
A JD (Job Description) is the industry's most honest vote on what knowledge it needs. Behind every requirement a company writes into a JD lies hard budget, a project schedule that's stuck, and the genuine pain of "we can't ship this without hiring this kind of person." Tech blogs can lie (that's marketing), conference talks can run ahead of reality (that's vision), but hiring demand can't — it has to describe concrete work that needs someone to do it right now. Every component covered earlier in this handbook — context engineering, the tool system, observability — has already been priced and validated by the hiring market: whether it's worth learning, and to what depth. What this module does is dig those answers out systematically.
About the data: how this survey was done
Every conclusion in this module is based on a focused search conducted August 11–12, 2026, covering 44 real, live job postings:
| Segment | Companies | Roles | Companies covered |
|---|---|---|---|
| International | 10 | 17 | Anthropic, OpenAI, Microsoft, Google DeepMind, Meta, Amazon AWS, xAI, Cognition, Cursor (Anysphere), Notion |
| Chinese | 12 | 27 | ByteDance, Tencent, Alibaba, Baidu, Meituan, Moonshot AI, Xiaohongshu, Ant Group, DeepSeek, MiniMax, Kuaishou, Zhipu (source note only) |
Sources fall into three confidence tiers; keep them apart as you read:
- Official (full text): complete JDs scraped directly from company career sites (Greenhouse, Ashby, amazon.jobs, jobs.bytedance.com, careers.tencent.com, job.xiaohongshu.com, etc.). High confidence — all citations defer to these.
- Official (title only): the career site confirms the role title, location, and team that's hiring, but the JD body couldn't be captured because it renders via JavaScript. Only title-level information is recorded; responsibilities are never filled in.
- Aggregator/search snapshots: third-party mirrors or search-engine snapshots from sites like BOSS Zhipin, Liepin, and Niuqi Zhipin. Medium-to-low confidence; every citation from these is explicitly flagged in this module.
A note on timeliness
Job postings get taken down. This survey is a snapshot of the hiring market in mid-August 2026: a well-known Anthropic FDE role was apparently delisted during the search window and couldn't be included. Read this module as a profile of industry demand at one specific moment, not as an eternal career guide. For the current status of individual posting links and the full original JD text, see the JD list.
A survey discipline
For any role where only the title could be captured and not the body, we mark it "unconfirmed" rather than inventing responsibilities. That's why some well-known companies (Zhipu, for one) appear only as a source note — gaps in the data are themselves honest information.
The role landscape: what international companies are hiring for
The 17 international roles fall into three clear tiers by title:
text
┌──────────────────────────── International Agent/LLM Engineering Role Landscape ─────────────────────────────┐
│ │
│ Customer delivery tier (hottest) │
│ Applied AI Engineer / Forward Deployed Engineer (FDE) │
│ Anthropic×2 · OpenAI · Microsoft · Cognition×2 · Notion │
│ Core focus: deploy models/products into customer environments; distill playbooks and reusable assets │
│ │
│ Harness core tier (literally shares its name with this handbook) │
│ Software Engineer, Agent Harness / Agent Evaluation │
│ Cursor (Anysphere) — agent loop, tools, guardrails, model routing │
│ │
│ Research & model tier │
│ Research Engineer / Applied Scientist │
│ OpenAI · DeepMind · Cognition Post-Training · Amazon×2 │
│ Core focus: post-training, RLHF, agentic data and algorithms; a master's or PhD is generally required │
└────────────────────────────────────────────────────────────────────────────────────────────────────────────────┘A few signals worth noting:
- FDE/Applied AI has become the biggest hiring category. Anthropic, OpenAI, Microsoft, Cognition, and Notion are all hiring this "embedded at the customer site" type of engineer, with strikingly consistent responsibilities: walking the customer from discovery to deployment, then distilling frontline patterns back into the product. The hard currency for these roles happens to be harness knowledge — Anthropic's JD (Applied AI Engineer, London) explicitly names prompting, context engineering, agent architectures, evaluation frameworks, and deployment at scale.
- Salaries are transparent, and the bands are wide. International official JDs routinely disclose pay ranges: Anthropic Applied AI (San Francisco/New York/Seattle) $200,000–$320,000; Microsoft FDE (Health) $142,800–$274,800 at IC5, up to $331,200 at IC6 in NYC; Cognition Applied AI $180,000–$225,000; Notion early-career roles $130,000–$150,000. Even Amazon's PhD-gated Applied Scientist (Agentic AI) has a base of $171,600–$222,200.
- Experience requirements cluster around 3–6 years, and varied backgrounds are accepted (FDE / SWE / technical PM / technical founder all count, per the Anthropic JD's own wording). Cursor's Agent Harness role lists no hard degree or years-of-experience gate at all — only that you have "built complex agentic products or infrastructure."
Lay out the official JDs that disclose pay, grouped by role type, and the band differences are obvious (all annual base, excluding bonus/equity):
| Role type | Representative role | Salary band |
|---|---|---|
| Customer delivery (FDE/Applied AI) | Anthropic Applied AI (San Francisco/New York/Seattle) | $200,000–$320,000 |
| Customer delivery (FDE/Applied AI) | Cognition Applied AI Engineer | $180,000–$225,000 |
| Customer delivery (FDE/Applied AI) | Notion FDE (Tokyo) | Not disclosed |
| Research/model (PhD track) | Amazon Applied Scientist II (Agentic AI) | $171,600–$222,200 |
| Research/model (PhD track) | Google DeepMind Research Engineer | $141,000–$202,000 |
| Early career (<2 years' experience) | Notion Software Engineer, Early Career (AI) | $130,000–$150,000 |
| Harness core | Cursor Software Engineer, Agent Harness | Not disclosed |
Two readings: first, there is a high-pay track that doesn't require a PhD — the top band sits at the customer-delivery layer, not the research layer, and engineering skill plus harness knowledge is itself hard currency. Second, the harness-core tier (Cursor) is the only one that doesn't disclose pay — benchmarked against engineer bands for the same company in the same location, the negotiation headroom is probably substantial.
The role landscape: what Chinese companies are hiring for
The 27 Chinese roles have a noticeably different shape:
- "Agent Harness" became an official job title outright. ByteDance (Agent Harness Engineer - AI Data & Security), Tencent (Hunyuan AI Agent Harness Engineer), Meituan (Agent Harness Engineer, owning the Tabbit Agent Harness), Xiaohongshu (Agent Harness Engineer), and DeepSeek (an Agent Harness team, live on its official recruiting page) — five companies hiring under this exact title. That is the most eye-catching change in the 2026 Chinese hiring market.
- Responsibilities converge to a striking degree. ByteDance's JD asks you to iterate on "agent planning, tool orchestration, RAG augmentation, long-context management, and the core task-scheduling pipeline," and to build a quantitative evaluation system with end-to-end observability; Tencent Hunyuan's Harness Engineer role likewise centers on tracing & observability, automated eval pipelines, A/B testing, and agent debugging tools. Runtime, evals, observability, and sandboxing are the standard four-piece kit on nearly every harness role.
- Hard requirements follow a familiar pattern: a bachelor's degree or above in a CS-related field, 2–5 years (ByteDance's harness role) or 3–5 years (Xiaohongshu's harness role) of backend/AI engineering experience; proficient Python, working familiarity with Go/Java; named frameworks such as LangChain/LangGraph.
- Pay disclosure runs opposite to the international market: it almost never happens. Chinese official JDs rarely state salary; the only data points come from aggregator snapshots (medium-to-low confidence): Ant Group's AReaL Agent Engineer at 35–65K RMB monthly on a 15-month pay scale (Liepin snapshot), MiniMax's Automation Testing Agent Developer at 30–50K (Jobui snapshot). Salary research before negotiating means finding other channels.
A new signal written into the hard requirements
The same shift shows up on both sides of the market: proficiency with AI coding tools is now a hard requirement. Tencent Hunyuan asks for "heavy programming with tools like Cursor / Claude Code / Codex, and first-hand feel for the capability limits and failure modes of agentic coding"; Alibaba's internship posting for the class of 2027 wants "power users of AI coding tools such as Cursor and Claude Code"; Notion's early-career role also expects you to track and use AI development tools. Only those who can use a harness have earned the right to build one.
Compare the two markets side by side and the differences cluster along four dimensions:
| Dimension | International | Chinese |
|---|---|---|
| Job naming | Applied AI Engineer / FDE is the mainstream; only Cursor explicitly uses "Agent Harness" | "Agent Harness Engineer" is now an official title at multiple major companies (ByteDance, Tencent, Meituan, Xiaohongshu, DeepSeek) |
| Salary disclosure | Official JDs routinely disclose bands ($130K–$330K annual base) | Official JDs almost never disclose; scattered data points only from aggregator snapshots (30–65K monthly base) |
| Degree & experience | Mostly 3–6 years; research roles want a PhD; engineering roles accept varied backgrounds; Cursor sets no hard gate | Generally a bachelor's degree or above plus 2–5 years' experience; clear internship and campus-hiring channels (Alibaba class of 2027, Baidu campus recruiting, Meituan rolling internships) |
| Skill emphasis | Customer delivery, evals, production deployment, cross-team collaboration | Runtime/execution engine, evaluation systems, observability, sandboxing — leaning more toward platform and infrastructure |
One reasonable reading: internationally, harness demand is bundled into delivery-oriented roles like FDE/Applied AI ("deploying agents into customer environments" is itself harness engineering), whereas Chinese companies hire for harness as a standalone trade. The paths differ, but the core-capability checklist is the same — which is exactly what the component chapters of this handbook cover.
Key finding: "Agent Harness" became a job title in both markets at once
This is the survey's most important finding, and it deserves a section of its own.
"Agent Harness" is no longer just a term from engineering blogs. It now appears in the official job titles on the career sites of leading companies in both markets:
| Company | Job title | Source |
|---|---|---|
| Cursor (Anysphere) | Software Engineer, Agent Harness | Official careers page |
| Tencent | Hunyuan AI Agent Harness Engineer | Official (full text) |
| ByteDance | Agent Harness Engineer - AI Data & Security | Official (full text) |
| Meituan | Agent Harness Engineer (Tabbit) | Official (listing-page excerpt) |
| Xiaohongshu | Agent Harness Engineer | Official (full text) |
| DeepSeek | Agent Harness team | Official (title only) |
And it goes beyond titles. Tencent Hunyuan's eval-infra role lists among its responsibilities "starting from the harness and the scoring logic, ensuring evaluation results stay accurate and trustworthy after the platform refactor"; Moonshot AI's multi-agent product engineer role requires that you "can quickly hand-roll your own harness to validate agent collaboration and orchestration efficiency," and calls the collaboration protocol itself "a new kind of harness." Word-frequency counts show that 10 of the 27 Chinese roles explicitly contain the word "Harness."
For readers of this handbook, that is direct validation: the stack you're learning — agent loop, context engineering, the tool system, evals, observability — is precisely the capability combination the hiring market has put an explicit price on, while supply remains scarce. The handbook's What Is an Agent Harness claims that "the harness determines the agent's ceiling"; the hiring market has a matching version of that claim: your harness engineering skill determines your market price.
Secondary finding: the model side and the engineering side have split into separate tracks
The 44 JDs also expose a clean layering of the hiring market — model-side roles and engineering-side roles have become two distinct career tracks:
- Model side: post-training, RL, agentic data, and evaluation benchmarks. On the Chinese side it concentrates at Tencent Hunyuan (sandbox/RL platform under Agent Infra), Baidu (SFT/RLHF), Ant Group (agent post-training), Kuaishou (end-to-end RL training of an agentic reasoning model), and DeepSeek; internationally at OpenAI Research Engineer, Cognition Post-Training, and Amazon Applied Scientist. A master's/PhD, published papers, or large-scale training experience is generally required.
- Engineering side: harness/infra/application delivery. On the Chinese side it concentrates at ByteDance, Meituan, Xiaohongshu, and Tencent's CSIG/WXG; internationally at the delivery and platform roles of Anthropic, Microsoft, Cognition, and Notion. What's wanted is solid engineering fundamentals plus harness knowledge, with far friendlier degree requirements.
The practical meaning of this split: harness is the track you can walk without a PhD. Chinese engineering-side roles typically gate on a bachelor's plus 2–5 years of experience; internationally, Cursor sets no hard degree or experience gate at all — what decides your competitiveness is verifiable engineering evidence such as "having built complex agentic products or infrastructure." That is also why this module's resume analysis makes demonstrable project evidence the focus of its checklist.
How to read this module
Three articles answer three questions; read them in order or dip in as needed:
| Article | Question it answers | Best for |
|---|---|---|
| The JD list | Who is hiring, for what, and at what pay? The complete list of all 44 roles: company, position, responsibilities, hard requirements, salary, source links, and confidence ratings | Anyone actively job hunting and ready to apply |
| The skill map | What capabilities do these roles actually demand? Word frequencies and requirements from the 44 JDs, aggregated into a demand map of harness knowledge points that maps one-to-one onto the handbook's component chapters | Anyone planning a learning path or closing knowledge gaps |
| The resume analysis | How far is my resume from these roles? Real JD requirements as the yardstick — checking your resume item by item for matches and gaps, with directions for fixing them | Anyone doing final polish before applying |
One recommended flow: get the full demand picture in the skill map first, go back to the handbook's core components chapter to fill in knowledge, use Build Your Own Harness to produce project evidence you can put on a resume, then pick targets against the JD list and calibrate with the resume analysis.
Further reading
- What Is an Agent Harness — the precise definition of what this title refers to in the hiring market
- Anatomy of a Harness — which component each JD staple (runtime, tools, context, evals) corresponds to
- Observability — one of the areas where harness-role responsibilities overlap most across both markets
- The Claude Code case study and the Cursor case study — complete teardowns of the product harnesses at two companies currently hiring harness engineers
- Build Your Own Harness — turn JD requirements into project evidence on your resume
- Common Pitfalls — the blind spots most likely to surface when an interviewer probes your harness design
- Glossary — quick reference for JD terms like FDE, evals, and agentic