The highest-signal line in DeepSeek's Harness hiring note is the formula: Model + Harness = Agent. The model sets the capability ceiling. The harness decides how that capability enters tasks, tools, context, evals, and product feedback loops.
The boundary
The job descriptions define harness as everything outside the model itself.
That moves the discussion from prompt engineering to the system layer. Prompt is an interface. Harness covers context management, long-term memory, tool APIs, subagents, multi-agent workflows, evals, annotation strategy, real-task feedback, user communities, product metrics, and feedback into model training.
An agent is model capability scheduled by a runtime and improved through a feedback system. That is why the Learn path starts with agent loops, tools, memory, and evals before moving into production and prompt-template tactics.
The three roles
| Role | Core question | Signal in the JD |
|---|---|---|
| Researcher | Where is the frontier? | Context, long-term memory, subagents, multi-agent systems, self-improving agents, benchmarks, annotation |
| Engineer | How does it become product? | Desktop agent product, architecture, toolchain, internal real tasks, developer experience, quality |
| PM | Does it actually help? | Roadmap, user need, community feedback, interviews, A/B tests, gradual rollout, statistics |
This goes beyond the old pipeline of research writes ideas, engineering implements, and product packages. The loop is tighter: real tasks produce traces and feedback; the harness team changes tools, context, evals, and product surfaces; the model team absorbs the training signal that survives.
The skill tree
1. Agent primitives
All three roles mention LLM APIs, KV cache, agent loops, tool use, reasoning, planning, skills, MCP, memory, subagents, and multi-agent systems. The baseline is understanding why an agent can keep acting across turns, tools, and observations. Start with Agent basics and Design patterns.
2. Context engineering
Context work is information scheduling: what appears, when it appears, how it is structured, who compresses it, when it enters long-term memory, and when it is discarded. Long tasks often fail here before they fail in the prompt. Claude Code's Context and CLAUDE.md, plus OpenClaw's Memory, are concrete entry points into this layer.
3. Tools and runtime
The engineering role names a desktop agent product. Desktop means files, apps, permissions, local state, recovery, context switching, and safety boundaries. Tool names, arguments, return shapes, error types, and permission policy all shape model planning. That is the practical value of
MCP and Skills: capabilities become composable, reviewable interfaces. For reference, compare Claude Code's MCP, Hooks, and Subagents with OpenClaw's Tools, Workspace, and Security.
4. Evaluation and data
Researchers own benchmarks and annotation. PMs own surveys, interviews, A/B tests, gradual rollout, and statistics. Engineers iterate from internal real tasks. Same point: agents need measured tuning. Evals make changes comparable; real tasks keep benchmarks grounded; product metrics keep research progress tied to user value. See Evaluation, Production, and Claude Code's Debugging.
5. Model-harness co-evolution
The descriptions repeatedly mention collaboration with the model training team. Harness becomes a feedback source and adaptation layer. Traditional benchmarks check final answers. Harness asks how the model behaves inside the loop: tool choice, plan correction after observation, subagent result integration, refusal behavior, and long-context task intent. Farther down the stack, this is where SFT for Agents and LLM Serving become relevant.
6. Model-behavior taste
Every role requires heavy agent product use. The PM JD lists Claude Code, Cowork, Codex, Cursor, OpenCode, GitHub Copilot, Manus, OpenClaw, and Hermes. The job calls for judgment: noticing sycophancy, hallucination, lazy tool use, premature stopping, false confidence, and context drift. Claude Code's Plan Mode and OpenClaw's Operations are useful places to observe those boundaries.
7. Agent-native execution
Researchers need fast prototypes. Engineers need quality in unfamiliar stacks. PMs are expected to vibe-code and prototype. The working mode is already agent-native. Harness work assumes people can direct agents. Claude Code's Workflows and Automation, plus OpenClaw's Multi-Agent, are closer to that working style than a prompt catalog.
For learners
The implied learning order is compact:
- Agent loop, tool use, memory, planning.
- MCP, skills, tool design, permission boundaries.
- Evals, traces, annotation data, real-task replay.
-
Subagents and multi-agent patterns
, task decomposition. - Production, cost, safety, user feedback, product metrics.
This separates two abilities. Using agents improves personal throughput. Directing agents requires systems judgment. DeepSeek Harness is hiring for the second.
Bottom line
The value of these JDs is that they name the capability model. The model layer still matters, but agent products increasingly compete outside the model: context organization, tool design, trace replay, eval definition, and feedback capture.
Prompt is the entry point. Harness is the compounding asset.