Agentic工程优于Vibe：DeepSeek Harness + Karpathy
Key Questions
Why did Andrej Karpathy join Anthropic?
Karpathy, known for coining 'Vibe Coding,' moved to Anthropic to advance agentic engineering approaches. This shift aligns with Anthropic's focus on structured AI workflows over purely intuitive coding methods.
What is agentic engineering compared to vibe coding?
Agentic engineering emphasizes systematic workflows, harnesses, and multi-step agent coordination for reliable results. Vibe coding relies more on natural language prompts for quick prototypes but often faces scalability issues.
What is DeepSeek's Harness team working on?
DeepSeek is building a Harness team to compete with Claude Code by developing terminal-based coding agents that handle file operations, shell commands, and Git management. An open-source project called DeepSeek-TUI has gained attention in developer communities.
How long and how much did Anthropic's agentic project take?
Anthropic completed a project with one-sentence requirements in 6 hours at a cost of $200 with zero human-written code. This demonstrates the potential of agentic systems to handle full project execution independently.
What were the results of Cursor testing mentioned?
Tests showed an 81% failure rate alongside successful outcomes in some cases. This highlights ongoing challenges in reliability for current AI coding tools.
How does vibe coding perform in complex projects?
Initial results can be promising for simple tasks, but complexity leads to accumulating issues like maintainability problems. Teams report needing to shift toward more structured agentic methods as projects grow.
What role does context play in AI coding tools?
Effective tools require strong context management to avoid quality drops when using the same model across different interfaces. Experiments show output differences stem more from tool design than the underlying model.
Are multiple agents being used in agentic workflows?
Yes, approaches like software factories involve teams of agents working together with correctness benchmarks. This is positioned as the next step beyond basic vibe coding.
Karpathy加入Anthropic提出'Agentic engineering'；DeepSeek组建Harness团队对标Claude Code，Agentic工作流主导；Anthropic 6小时+、200美元实践；Cursor测试与81%失败率并存。