Back to home

Compare

Comparing: Coding Agents Close the Software Engineering Loop, Paving the Way for AI Automation & Coding Agent 跑通软件工程闭环,AI 自动化先从代码切入

AEN
Coding AgentGoogleOpenAI·

Coding Agents Close the Software Engineering Loop, Paving the Way for AI Automation

Codex, Claude Code, Kimi Code, and other Coding Agents—AI assistants capable of autonomously writing, modifying, and validating code—are becoming core products for foundation-model companies, with Google making particularly notable progress over the past two months. We see this as more than an upgrade to programming tools: it is a critical entry point for AI’s shift from “answering questions” to “completing digital tasks.”

What this is

The author regards Coding Agents as an important step toward AGI, based on three arguments. First, software is the foundation of digital automation: the information and actions people accumulate on computers generally must be invoked through code to be executed at scale. Second, open-source communities such as GitHub provide enormous, low-cost corpora for training and validation. Third, code execution results can be checked quickly, making software a better fit for establishing feedback loops than long-cycle decisions such as drug discovery or medical treatment plans.

Coding Agents therefore do more than increase coding speed. They enable models to understand objectives, modify files, run tests, and fix errors. Google Antigravity, OpenAI Codex, Claude Code, and the related Agent products from Kimi Code and DeepSeek are all competing for control of the workplace entry point along this path.

Industry view

We agree that code is a high-quality starting point for AI automation: outcomes are verifiable, processes are reproducible, and a successful experience can be turned into a reusable tool. But the claim that “Coding Agents are an important step toward AGI” does not mean they can reach AGI alone. Software environments are highly standardized, while enterprise workflows are full of permissions, legacy systems, and accountability boundaries. Code passing tests does not mean the business outcome is correct.

The risks are equally clear. Once autonomously executing Agents have access to terminals, code repositories, and credentials, mistakes can compound rapidly. Enterprises also face challenges involving audits, data leakage, and responsibility attribution. The original article’s assessment of Google is based on recent product experience and improvements, which demonstrates the pace of competition but does not prove that Google has established a durable advantage. Interaction design and industry experience across today’s Agent products remain immature, and a single trial cannot replace metrics such as retention, task success rate, and cost.

Impact on regular people

For enterprise IT: Codebases, testing systems, and access controls will become foundational requirements for deploying Agents. In the near term, value is more likely to emerge in testing, migration, operations, and internal-tool development—not in unsupervised changes to critical systems.

For individual careers: People who can write code will find it easier to delegate repetitive work to Agents, while product, operations, and business staff will increasingly participate in development through natural language. Competitiveness will depend more on the ability to define objectives, validate results, and take final responsibility.

For consumer markets: As software-development costs continue to fall, standalone tools and niche applications will proliferate, putting pressure on proprietary commercial software pricing and subscription models. At the same time, as intelligent execution enters sensitive areas such as payments and account management, permission prompts and human confirmation will not disappear.

Source: juejin.cn
BZH
Coding AgentGoogleOpenAI·

Coding Agent 跑通软件工程闭环,AI 自动化先从代码切入

Codex、Claude Code、Kimi Code 等 Coding Agent(能自主编写、修改并验证代码的 AI 助手)正在成为大模型公司的核心产品,Google 近两个月的进展尤其明显。我们认为,这不只是编程工具升级,而是 AI 从“回答问题”走向“完成数字任务”的关键切入口。

这是什么

作者把 Coding Agent 视为通往 AGI(通用人工智能)的重要环节,理由集中在三点:第一,软件是数字自动化的基础,人类在电脑上沉淀的信息和操作,要批量执行往往都要通过代码调用;第二,GitHub 等开源社区提供了规模庞大且成本较低的训练与验证语料;第三,代码运行结果可以较快检查,比药物研发、医学方案等长周期决策更适合建立反馈闭环。

因此,Coding Agent 并非单纯提高写代码速度,而是让模型能够理解目标、修改文件、运行测试并修复错误。Google Antigravity、OpenAI Codex、Claude Code,以及 Kimi Code、DeepSeek 的相关 Agent 产品,都沿这条路线争夺工作入口。

行业怎么看

我们认同代码是 AI 自动化的高质量起点:结果可验证、过程可复现,也能把一次成功经验转化为可重复调用的工具。但“Coding Agent 是通往 AGI 的重要一环”不等于它能单独通向 AGI。软件环境高度标准化,而企业流程充满权限、遗留系统和责任边界;代码通过测试,也不代表业务结果正确。

风险也很明确。自主执行的 Agent 一旦拥有终端、代码库和凭证权限,错误可能被连续放大;企业还面临审计、数据泄露和责任归属问题。原文对 Google 的评价基于近期体验和产品进步,这可以说明竞争节奏,却不足以证明其已经建立稳定优势。当前各家 Agent 的交互设计和行业积累都处于早期,单次体验不能替代留存率、任务成功率和成本等指标。

对普通人的影响

对企业 IT:代码库、测试体系和权限控制将成为部署 Agent 的基础条件。短期价值更可能出现在测试、迁移、运维和内部工具开发,而不是无人监督地改动关键系统。

对个人职场:会写代码的人更容易把重复工作交给 Agent,但产品、运营和业务人员也将通过自然语言参与开发。竞争力将更多取决于描述目标、验证结果和承担最终责任的能力。

对消费市场:当软件制作成本继续下降,独立工具和小众应用会增多,闭源商业软件的价格与订阅模式可能受到压力。同时,智能执行进入支付、账号等敏感环节后,权限提示和人工确认不会消失。

Source: juejin.cn