跪拜 Guibai
← All articles
Frontend · DeepSeek · AI Programming

DeepSeek Harness Is Not a Claude Code Killer — It’s a Composable Agent Runtime

By kyriewen ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Harness shifts the AI coding tool conversation from “which assistant writes better code” to “who controls the runtime.” Its plugin architecture lets teams build custom agent loops, security pipelines, and tool orchestrations that closed assistants cannot replicate, but the v0.1 maturity means it is not yet a practical replacement for Claude Code in daily work.

Summary

DeepSeek Harness launched as open source under MIT and immediately drew comparisons to Claude Code, but the two tools occupy different categories. Harness is a Web UI-based agent runtime built on a Cordis microkernel where every component — model provider, tools, session strategy, sandbox, and the agent loop — is a replaceable plugin. It supports nearly 40 model providers and introduces a PTC mode for deterministic TypeScript-based tool orchestration. Claude Code, by contrast, is a terminal-native, model-locked assistant with a deeply optimized, non-replaceable agent loop and mature self-correction that automatically runs type checks and verifies rendering. In practice, Harness v0.1 lacks that verification loop and requires manual patching to enable Skills, making it a developer preview for Harness authors rather than a daily driver for working programmers.

Takeaways
— Harness is a Web UI agent runtime, not a terminal-native coding assistant; it runs on localhost:3080 and requires port management.
— Skills are disabled by default in the Web version with no error — a manual YAML patch is needed to enable them.
— Four built-in modes are really preset plugin combinations: Standard, PTC (programmatic TypeScript orchestration), Minimal, and Creative.
— PTC mode lets developers pre-define deterministic multi-step tool call sequences in TypeScript that pass through the full security pipeline.
— The architecture treats the agent loop itself as a replaceable plugin, unlike Claude Code where the loop is fixed and deeply optimized.
— Every tool call passes through a pipeline of Hook → Approval → Permission Check → Sandbox → Timeout Control, with developer-injectable plugins at each stage.
— Harness supports nearly 40 model providers and can run Claude models inside its own harness, enabling controlled A/B experiments.
— In a React component task, Harness generated code but did not run type checks or verify rendering; Claude Code automatically ran tsc, fixed errors, and started a dev server.
— DeepSeek simultaneously raised V4-Pro model pricing, signaling a strategy where the free open-source Harness ecosystem drives paid model usage.
— The formula Model + Harness = Agent defines how future AI programming tools will be selected and combined, more than any single product.
Conclusions

Calling Harness a Claude Code competitor misses the point: it competes at the framework layer, not the assistant layer, and its real rivals are agent-building platforms like LangChain or CrewAI.

The decision to ship as a browser-based Web UI rather than a terminal tool is a deliberate architectural bet — it decouples the UI from the runtime so the UI itself becomes a replaceable plugin, but it sacrifices the low-friction terminal experience that makes Claude Code feel immediate.

Harness’s PTC mode is a genuinely novel idea: giving developers a TypeScript API to pre-orchestrate tool calls flips the agent paradigm from model-driven autonomy to developer-defined determinism, which matters for safety-critical or compliance-heavy workflows.

The tool-call pipeline with injectable hooks at every stage is a more granular security model than Claude Code’s hook system, which can intercept but cannot inspect parameters or rewrite results.

DeepSeek’s simultaneous price hike on V4-Pro alongside the Harness open-source release mirrors the OpenAI playbook — give away the platform, charge for the compute — and suggests the company sees Harness as a model-distribution channel, not a standalone product.

The 4,000+ existing Cordis plugins give Harness a head start on ecosystem, but the gap between a plugin marketplace and a polished agent loop with self-correction is where the real engineering work lies.

Concepts & terms
Harness
In DeepSeek’s architecture, a Harness is the configurable runtime layer that sits between a model and the outside world — it controls context management, tool registration, session strategy, error retry logic, and termination conditions. The same model in different Harnesses can behave very differently.
Cordis microkernel
The minimal core of DeepSeek Harness that only handles plugin loading, unloading, and dependency resolution. All other functionality — including the agent loop, UI, and tools — is implemented as plugins on top of this kernel.
PTC Mode (Programmatic Tool Calling)
A Harness mode that lets developers write TypeScript code to pre-define deterministic sequences of tool calls, which then pass through the full security pipeline. It contrasts with letting the model autonomously decide which tools to call and when.
Agent Loop
The core execution cycle of an AI coding agent — observe, plan, act, verify, repeat. In Harness, this loop is a replaceable plugin; in Claude Code, it is a fixed, deeply optimized internal component that includes automatic type checking and rendering verification.
From the discussion

The discussion splits between those who see Claude Code's skills ecosystem as the real competitive advantage and those who defend DeepSeek's plugin architecture and cache-hit design as promising. A sharp disagreement emerges over DeepSeek's actual coding competence: one side calls it unusably dumb, while another points to speed and cost-performance as redeeming factors. One comment dismisses the article itself as AI-generated fluff.

— Claude Code's moat is its skills ecosystem, not the agent runtime itself.
— DeepSeek's plugin philosophy and cache-hit mechanism are technically impressive and improving rapidly.
— DeepSeek's coding ability in practice is poor, especially when combined with Claude Code.
— DeepSeek offers fast inference and strong cost-performance, which offsets some capability gaps.
— The article reads like AI-generated content, undermining its credibility.
Featured comments
咬代码的兽

Pretty fair comparison. Claude Code's real moat right now is still the skills ecosystem — for example, the recently popular gc-minimal-zine-poster skill, which can produce a minimalist magazine-style poster in one sentence. There are basically no Chinese tutorials at the moment; I've put one together: installation commands, two installation methods for v0-1 and v0.3.1, five modes, and prompt examples https://xia345.com/post/post-1785399726870 Also recommending an AI Agent community xia345.com — it aggregates skill tutorials, LLM rankings, and various Coding Plan news. Worth bookmarking if you follow the Agent ecosystem.

用户32408878106

For DeepSeek, I think its plugin philosophy and cache-hit mechanism are really impressive. Saw a guru online yesterday who spent a long time tuning it, and it's already much better than before — test blocks are catching up to top-tier levels. The dsh-routing-suite plugin — this guru is now debugging the Pro model.

kyriewen  → 用户32408878106

Very likely to become the mainstream in the future.

用户57505257976

Anyone who's actually written code with DS knows how stupid this thing is, especially paired with CC — it's just unbelievably dumb.

kyriewen

It's okay — fast speed, pretty high cost-performance ratio.

我是大大大佬

Reeks of AI-generated writing.

kyriewen

[flash of inspiration] Keep it up, you write too

See top comments, translated →
Source: juejin.cn ↗ Google Translate ↗ Backup ↗