跪拜 Guibai
← All articles
AI Programming · Artificial Intelligence · Agent

Why Every AI Lab Is Betting on Coding Agents to Reach AGI

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

If coding agents are the most direct path to AGI, then every developer's daily tooling is the testbed for the next generation of AI. Companies are betting their futures on agent quality, and that will show up as better, faster coding assistants long before AGI arrives.

Summary

Coding agents have become the center of the AI industry. Codex, Claude Code, Kimi Code, and Grok are all examples. Three structural advantages explain why. First, the digital world already runs on software, so any automation goal ultimately requires programming; AI needs automation before intelligence. Second, the open-source corpus is unmatched: GitHub has held mature code for a decade, and programmers are the one community that freely publishes their best work, giving AI the cheapest, highest-quality training data. Third, software engineering offers the fastest verification loop of any domain: you can check for bugs almost immediately, unlike medicine or drug design where validation takes years.

The argument goes further: AGI cannot be reached in a single leap. Coding agents are a subset of AGI, but the most important one, and arguably its basic infrastructure. Anthropic committed to this path early and is succeeding, while OpenAI and Google continue AGI research alongside their coding agent efforts. Google now covers coding, images, video, music, and world models, but its Antigravity agent has the worst UX of any agent the author has tried, while the new Gemini 4 Carbon model feels strong. DeepSeek's agent also has poor UX today.

The deeper point is that agent product design has no legacy. AI agents only truly appeared this year, so every company is starting from zero. Judging them by their current state is a mistake; relative progress over the last two months is what matters, and Google is moving fastest.

Takeaways
— Codex, Claude Code, Kimi Code, and Grok are all current examples of coding agents.
— Programming is a prerequisite for automating any digital task, so AI needs automation before intelligence.
— Open-source code from places like GitHub is the lowest-cost, highest-quality training data for AI because programmers are unusually willing to share.
— Software engineering allows near-instant verification of AI output, unlike medicine or drug development where validation takes years.
— AGI cannot be achieved in one step; coding agents are a subset and arguably the most important piece.
— Anthropic focused on coding agents early and is succeeding, while OpenAI and Google pursue both coding agents and AGI research.
— Google's product line spans coding, images, video, music, and world models, but its Antigravity agent has the worst UX among agents tested.
— Gemini 4 Carbon is a strong new model, and Google has shown the fastest relative progress in the last two months.
— DeepSeek's agent currently has poor UX, but agent architecture is so new that present-state judgments are premature.
Conclusions

The article's claim that software engineering's fast verification loop is a unique advantage implies that domains like medicine or law will lag behind in agent adoption because they lack that feedback cycle.

The dismissal of current UX as a criterion for future judgment is a reminder that agent design is only a year old; the market is still in a pre-Chrome moment where the eventual winner could come from anywhere.

The author's appreciation for Google's relative progress despite its poor agent UX suggests model capability and breadth may outrank interface polish in this phase, reversing typical product wisdom.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗