SourcePandaily•Stalecollected in 62m
DeepSeek building coding agent team to rival Claude Code
DeepSeek is entering the coding agent race, potentially disrupting the market dominance of Claude Code and Copilot.
30-Second TL;DR
What Changed
DeepSeek is establishing a dedicated 'Harness' team for coding agents
Why It Matters
This signals a shift in the AI landscape where model labs are moving beyond chat interfaces into autonomous developer tools. It increases pressure on incumbents like Anthropic and GitHub Copilot to innovate faster.
What To Do Next
Monitor DeepSeek's GitHub repository and career page for the release of their agentic framework or open-source coding tools.
Who should care:Developers & AI Engineers
Key Points
- •DeepSeek is establishing a dedicated 'Harness' team for coding agents
- •The project is a direct response to Anthropic's Claude Code
- •New roles are currently open in Beijing to support this development
Key numbers87%13%
Deep Insight
Background and context from public sources — not the original article. 20 sources cited.
Enhanced Key Takeaways
- •DeepSeek's new coding agent project is tentatively named "DeepSeek Code" and the "Harness" team's focus is on developing components beyond the core model, such as tool use, planning, and memory.
- •The "Harness" team is specifically tasked with developing a "DeepSeek Desktop Agent" product, aiming to integrate real-world code testing feedback directly into the underlying model's development cycle.
- •DeepSeek has strategically recruited Cui Tianyi, a former Jane Street engineer, to its AI "harness" team, underscoring the company's commitment to building robust software infrastructure for autonomous AI agents.
- •DeepSeek's existing DeepSeek Coder models are trained on a massive dataset of 2 trillion tokens, consisting of 87% code and 13% natural language in both English and Chinese, and are available in various sizes from 1.3B to 33B parameters.
- •Anthropic's Claude Code, the direct competitor, launched as a research preview in February 2025 and achieved general availability in May 2025, evolving into a sophisticated multi-agent development platform with features like subagents and browser automation.
Competitor Analysis
Core Offering
- DeepSeek (Harness/DeepSeek Code - anticipated)
- AI Coding Agent (Desktop Agent focus)
- Anthropic Claude Code
- Agent-first coding tool (terminal, web, multi-agent platform)
- OpenAI Codex/Copilot
- AI coding assistant (IDE integration, CLI, multi-model backends)
- Cursor
- AI coding IDE (VS Code fork, agentic workflows)
Key Differentiators
- DeepSeek (Harness/DeepSeek Code - anticipated)
- "Model + Harness = Agent" approach, focus on tool use, planning, memory, and real-world feedback loops
- Anthropic Claude Code
- Multi-agent teams, browser automation, computer use, remote control, mobile continuity, high SWE-bench scores
- OpenAI Codex/Copilot
- Broad adoption, multi-model platform (Claude, Codex backends), CLI with specialized sub-agents, autopilot mode
- Cursor
- Market-leading AI coding IDE, project context across files, fast and accurate tab completions
Pricing
- DeepSeek (Harness/DeepSeek Code - anticipated)
- Not yet announced for "DeepSeek Code"
- Anthropic Claude Code
- Pro and Max subscribers ($20/month and $100–200/month respectively)
- OpenAI Codex/Copilot
- Typically subscription-based (e.g., GitHub Copilot $10/month)
- Cursor
- Subscription-based, annual recurring revenue over $500M
Benchmarks (DeepSeek Coder vs. others)
- DeepSeek (Harness/DeepSeek Code - anticipated)
- DeepSeek-Coder-Base-33B outperforms CodeLlama-34B by 7.9% (HumanEval Python), 9.3% (HumanEval Multilingual), 10.8% (MBPP), 5.9% (DS-1000). DeepSeek-Coder-Instruct-33B outperforms GPT-3.5-turbo on HumanEval.
- Anthropic Claude Code
- Opus 4.6 scores 80.8% on SWE-bench Verified, 55.4% on SWE-bench Pro.
- OpenAI Codex/Copilot
- GPT-5.5 reaches 82.7% on Terminal-Bench 2.0, outpacing Claude Opus 4.7 in OpenAI's published comparison.
- Cursor
- Composer 2.5 matches Opus 4.7 and GPT-5.5 benchmarks.
Availability
- DeepSeek (Harness/DeepSeek Code - anticipated)
- Under development, new team formed
- Anthropic Claude Code
- Web, terminal (CLI), iOS application
- OpenAI Codex/Copilot
- VS Code, JetBrains, Neovim, GitHub.com, CLI
- Cursor
- VS Code fork, IDE
| Feature/Aspect | DeepSeek (Harness/DeepSeek Code - anticipated) | Anthropic Claude Code | OpenAI Codex/Copilot | Cursor |
|---|---|---|---|---|
| Core Offering | AI Coding Agent (Desktop Agent focus) | Agent-first coding tool (terminal, web, multi-agent platform) | AI coding assistant (IDE integration, CLI, multi-model backends) | AI coding IDE (VS Code fork, agentic workflows) |
| Key Differentiators | "Model + Harness = Agent" approach, focus on tool use, planning, memory, and real-world feedback loops | Multi-agent teams, browser automation, computer use, remote control, mobile continuity, high SWE-bench scores | Broad adoption, multi-model platform (Claude, Codex backends), CLI with specialized sub-agents, autopilot mode | Market-leading AI coding IDE, project context across files, fast and accurate tab completions |
| Pricing | Not yet announced for "DeepSeek Code" | Pro and Max subscribers ($20/month and $100–200/month respectively) | Typically subscription-based (e.g., GitHub Copilot $10/month) | Subscription-based, annual recurring revenue over $500M |
| Benchmarks (DeepSeek Coder vs. others) | DeepSeek-Coder-Base-33B outperforms CodeLlama-34B by 7.9% (HumanEval Python), 9.3% (HumanEval Multilingual), 10.8% (MBPP), 5.9% (DS-1000). DeepSeek-Coder-Instruct-33B outperforms GPT-3.5-turbo on HumanEval. | Opus 4.6 scores 80.8% on SWE-bench Verified, 55.4% on SWE-bench Pro. | GPT-5.5 reaches 82.7% on Terminal-Bench 2.0, outpacing Claude Opus 4.7 in OpenAI's published comparison. | Composer 2.5 matches Opus 4.7 and GPT-5.5 benchmarks. |
| Availability | Under development, new team formed | Web, terminal (CLI), iOS application | VS Code, JetBrains, Neovim, GitHub.com, CLI | VS Code fork, IDE |
Technical Deep Dive
- DeepSeek Coder Models: Comprise a series of code language models trained from scratch on 2 trillion tokens.
- Training Data Composition: 87% code and 13% natural language (English and Chinese).
- Model Sizes: Available in various sizes, including 1.3B, 5.7B, 6.7B, and 33B parameters.
- Pre-training: Utilizes a repo-level code corpus with a 16K window size and an extra fill-in-the-blank task to support project-level code completion and infilling.
- Instruction Tuning: Base models are further fine-tuned with 2B tokens of instruction data to create instruction-tuned models (e.g., DeepSeek-Coder-Instruct).
- DeepSeek-R1 Architecture: Employs a "mixture-of-experts" architecture, allowing it to activate only a small number of parameters for a given task, enhancing efficiency and reducing costs.
- Training Techniques: DeepSeek-R1 incorporated thousands of "cold-start" data points to fine-tune the V3-Base model before applying reinforcement learning.
- Harness Concept: Defined as everything beyond the model itself, including context management, tool invocation, file reading/writing, terminal execution, and test feedback.
Future ImplicationsAI analysis grounded in cited sources
The AI coding agent market will see intensified competition, particularly from Chinese AI firms.
DeepSeek's direct move to rival Claude Code with a dedicated team indicates a strategic push from Chinese companies into advanced AI agent development, challenging established Western players.
The development of 'harness' infrastructure will become a critical differentiator for AI agents.
DeepSeek's emphasis on the 'Harness' (tool use, planning, memory, feedback loops) highlights a market shift where the surrounding engineering and operational framework is as crucial as the underlying AI model.
AI coding agents will increasingly move towards more autonomous and integrated development workflows.
The focus on desktop agents, real-world code testing feedback, and the evolution of competitors like Claude Code towards multi-agent platforms suggest a trend where AI agents will handle more complex, end-to-end development tasks with less human intervention.
Timeline
2023-05
DeepSeek founded by Liang Wenfeng.
2023-11
DeepSeek Coder, an open-source model for coding tasks, released.
2024-05
DeepSeek-V2 chatbot model released, gaining popularity in China for cost-efficiency.
2025-01
DeepSeek-R1 reasoning model and mobile chatbot application released, becoming a top downloaded app.
2025-02
Anthropic's Claude Code launched as a research preview.
2025-05
Claude Code became generally available alongside Claude 4.
2026-05
DeepSeek announced the formation of its 'Harness' team to develop AI coding agents.
- 2023-05DeepSeek founded by Liang Wenfeng.
- 2023-11DeepSeek Coder, an open-source model for coding tasks, released.
- 2024-05DeepSeek-V2 chatbot model released, gaining popularity in China for cost-efficiency.
- 2025-01DeepSeek-R1 reasoning model and mobile chatbot application released, becoming a top downloaded app.
- 2025-02Anthropic's Claude Code launched as a research preview.
- 2025-05Claude Code became generally available alongside Claude 4.
- 2026-05DeepSeek announced the formation of its 'Harness' team to develop AI coding agents.
Sources (20)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
1the-decoder.comvertexaisearch.cloud.google.com2kucoin.comvertexaisearch.cloud.google.com3phemex.comvertexaisearch.cloud.google.com4scmp.comvertexaisearch.cloud.google.com5github.iovertexaisearch.cloud.google.com6github.comvertexaisearch.cloud.google.com7huggingface.covertexaisearch.cloud.google.com8medium.comvertexaisearch.cloud.google.com9wikipedia.orgvertexaisearch.cloud.google.com10medium.comvertexaisearch.cloud.google.com11morphllm.comvertexaisearch.cloud.google.com12codegen.comvertexaisearch.cloud.google.com13mightybot.aivertexaisearch.cloud.google.com14builder.iovertexaisearch.cloud.google.com15mlq.aivertexaisearch.cloud.google.com16pandaily.comvertexaisearch.cloud.google.com17readthedocs.iovertexaisearch.cloud.google.com18mit.eduvertexaisearch.cloud.google.com19bentoml.comvertexaisearch.cloud.google.com20medium.comvertexaisearch.cloud.google.com
Weekly AI Recap
Read this week's curated digest of top AI events →
Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: Pandaily ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.