DeepSeek · Agent · AI Programming

DeepSeek's 36 New Roles Show Agent Engineering Is Now a Baseline Skill, Not a Specialty

By 沉默王二 · Jun 29, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

When a company with tens of millions of DAU makes Agent skills a requirement for 80% of its technical roles—including QA and mobile—it signals that the industry is standardizing around a new baseline. Developers who treat Agent engineering as optional risk being locked out of the most heavily funded teams.

Summary

A wave of hiring at DeepSeek reveals a hard shift in engineering requirements: Agent development has moved from a niche research topic to a mandatory skill across backend, frontend, testing, and product roles. The company plans to double every department's headcount, with 121 positions listed on recruitment platforms. The job descriptions demand fluency in LLM APIs, KV Cache, MCP, tool-use loops, and prompt engineering, even for traditional roles like QA and mobile development.

Two entirely new positions—Agent Harness and Agent Infra—formalize work that barely existed two years ago. Harness engineers build the orchestration layer around models (context, memory, tool calling, multi-agent coordination), while Infra engineers design cloud platforms that spin up millions of isolated sandboxes for Agent code execution. The JD for Harness explicitly values the ability to enter unfamiliar technical domains using AI assistance over expertise in any single language.

The hiring blueprint doubles as a career roadmap. Foundational computer science remains non-negotiable, but a third layer of AI-native skills now sits on top: understanding tokenization and context windows, building streaming services with SSE and WebSocket, and designing evaluation systems that replace deterministic tests with probabilistic benchmarks like HumanEval and SWE-bench.

Takeaways

— DeepSeek's parent company closed a 51 billion RMB (≈$7B) Series A, triggering a hiring push across 36 official roles and 121 platform listings.

— 80% of the job descriptions explicitly require Agent-related skills, including LLM APIs, KV Cache, MCP, tool use, and Agent loops.

— All departments will at least double in size as the technology evolves.

— Backend roles now demand streaming architectures (SSE, WebSocket, gRPC streaming) and sandbox infrastructure (gVisor, Firecracker) to support long-running, tool-calling Agent tasks.

— Frontend roles shift from chat-dialog UIs to Agent state visualization: rendering multi-step, interruptible execution flows with real-time status indicators.

— Testing roles replace deterministic pass/fail checks with probabilistic evaluation systems using benchmarks like HumanEval, MMLU, and SWE-bench, and require Go or Rust for performance-sensitive test infrastructure.

— Agent Harness is a new role that builds the orchestration layer around models—context management, memory, tool calling, sub-agent coordination, and self-evolution.

— Agent Infra is a new role that builds cloud platforms hosting millions of concurrent, millisecond-startup sandboxes for Agent code execution.

— Product manager roles now require Vibe Coding: the ability to use AI tools to build prototypes rather than just writing PRDs.

— The hiring criteria form a three-layer capability stack: computer science fundamentals, engineering skills (containers, CI/CD, streaming), and AI-native skills (LLM mechanics, Agent tooling, AI-assisted development).

Conclusions

Agent skills have crossed the chasm from research specialization to general engineering requirement faster than any previous paradigm—containers and cloud-native took years; Agent expectations compressed into roughly 18 months.

The explicit demand for Vibe Coding in product roles redefines the PM as a builder, not a spec-writer, and collapses the prototype-to-production handoff.

Requiring Go or Rust for testing roles signals that AI evaluation pipelines are now performance-sensitive infrastructure problems, not scriptable QA tasks.

DeepSeek's Harness JD doesn't ask for Rust expertise—it asks for the ability to learn Rust in a day with AI assistance. This inverts the traditional hiring signal from 'what you know' to 'how fast you can learn with AI.'

The 'use Agent to test Agent' pattern creates a compounding automation loop that could make manual QA for AI systems obsolete within two years.

Doubling every department while simultaneously redefining every role's skill set is a high-risk organizational bet that most Western companies would stagger across years, not attempt in a single hiring cycle.

Concepts & terms

Agent Harness

The orchestration layer that wraps a large language model to turn it into an autonomous agent. It handles context management, long-term memory, tool calling, sub-agent coordination, task planning, and self-evolution mechanisms—everything the model doesn't do natively.

Agent Infra

The cloud infrastructure purpose-built for hosting agent execution environments. It manages millions of concurrent sandboxes that start in milliseconds, enforce strict resource and network isolation, and safely execute arbitrary code on behalf of users.

ReAct Loop

A reasoning-and-acting pattern where an agent receives an instruction, decomposes it into steps, selects and calls external tools, observes the results, and decides the next action. Each cycle is a nested request-response loop that backend infrastructure must support.

Vibe Coding

Using AI coding assistants to build functional software prototypes through natural-language prompting, without deep expertise in the underlying languages or frameworks. The term emphasizes speed and experimentation over traditional software engineering rigor.

KV Cache

A key-value cache that stores intermediate attention states during large language model inference, avoiding recomputation of previous tokens. It is critical for performance in long-context and multi-turn agent interactions, and spans inference, storage, and agent system design.

MCP (Model Context Protocol)

A protocol for standardizing how AI models connect to external tools and data sources. It defines a common interface for tool use, allowing agents to call APIs, search databases, and execute code through a uniform mechanism.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗