What kind of “Lassoing Horse” is DeepSeek looking for?

On May 18th, DeepSeek posted a JD for “Agent Harness R&D Engineer.” Reading through it line by line, the profile becomes clear.

I. “2+ years of experience”

At the same level in big tech companies, the minimum requirement is 3-5 years. DeepSeek writes 2 years — which means they don’t want someone who coasts on tenure.

“Proficient in using AI Agent tools to write code”

Not a plus — it’s a requirement.

It’s followed by: “Able to write quality code even in languages and frameworks you’ve never encountered before.”

Meaning: Learn a new language in an hour with Claude Code, and still write production-ready code.

Most domestic developers’ understanding of AI-assisted coding is still stuck at “Copilot helps me auto-complete a line.” DeepSeek takes it straight to “I and the Agent finish the entire project together.”

II. “Deep experience with various Agent products”

The PM’s JD from the day before listed product names — Claude Code, Cursor, Cowork, Manus. This one didn’t list names; it wrote “heavy user” plus “integrated into work and life.”

What counts as heavy user? My standard:

• Using it at least 4 hours every day
• Can’t work without an Agent
• Has strong opinions on which products are good or bad

III. “Understands LLM and Agent principles”

Checklist:

• LLM API, KV Cache — fundamentals
• Agent Loop, Tool Use, Reasoning, Planning — the four pillars
• Skills, MCP, Memory — core of the Anthropic ecosystem, practically saying “we’re copying Claude Code”
• Subagent, Multi-Agent
• Prompt / Context / Harness Engineering

IV. The bonus points tell the real story

AI industry experience, fast product development in small teams, open-source contributions, working with researchers, English fluency to hang out in international communities, ACM awards, published papers, and the last one — “other extraordinary skills.”

Tying it all together: DeepSeek doesn’t want an engineer. They want a “researcher + engineer + community manager” triple threat.

Someone who can lasso a horse, write a research paper about horses, write reports in English for the international equestrian association, and ideally sing a couple of horse-themed songs around a campfire.

How many such people are there in China?

I estimate no more than 200.

Most are in the LLM teams at ByteDance, Alibaba, Tencent, or among Chinese engineers at Anthropic or OpenAI.

DeepSeek knows this, so the JD is extremely specific — the right people will see it and think, “This was written for me,” and come on their own.

V. The best part is the last line

“Other extraordinary skills related to this job. “

It says:

We don’t know what the final Harness will look like. If you have skills we haven’t listed, they might be exactly what we need — but haven’t realized we need yet.

The grassland is vast, the horses are wild. The method for lassoing them might not have been invented yet.

If you’re the one who can invent a new way to lasso, DeepSeek is waiting for you.