Follow ZDNET: Add us as a preferred source on Google.
ZDNET’s key takeaways
- Replit unveiled Agent 3 on Wednesday.
- Code-generation is one of the few viable business use cases for AI.
- However, Replit recently deleted a user’s entire codebase.
On Wednesday, AI startup Replit released Agent 3, an autonomous code generation system designed to help non-programmers with software development projects. It’s the latest in the industry-wide investment in vibe coding tools, one of the most popular applications of AI.
Also: I asked AI to modify mission-critical code, and what happened next haunts me
After receiving high-level instructions from a human user, Agent 3 is able to iteratively test and fix its own code. It can also build other agents, which can then be integrated into third-party apps like Slack, Telegram, or Gmail.
“The net outcome is an increase in your efficiency and productivity,” Replit wrote in the announcement, in keeping with what has been the dominant argument for both coding and non-coding AI tools.
The new agent, Replit added, is “up to 3x faster and 10x more cost-effective than Computer Use models,” referring to AI systems that can use computers the way humans can.
The rise of vibe coding
Vibe coding systems like Agent 3, which autonomously create software from natural language prompts, have become a key area of investment in Silicon Valley amid the ongoing AI boom. Code-generation is one of the few applications of generative AI systems, including agents, that has proven to be demonstrably useful for enterprises. Similar agents include Google’s Jules and OpenAI’s Codex (which OpenAI says it uses in-house).
Also: I spent $20 on Codex and got 24 days of coding work done in 6 hours – but there’s a big catch
Replit announced Wednesday that it had raised $250 million in its latest funding round, bringing its total valuation to $3 billion. Cursor and Cognition, two other AI startups specializing in code generation, are now valued at $10 billion and $10.2 billion, respectively, according to Reuters. And Anthropic, the company behind Claude, which has become the preferred AI coding assistant for many programmers, just announced that it’s now valued at a staggering $183 billion.
Many tech companies have begun using AI systems for low-stakes, routine coding tasks that have historically been delegated to younger and less experienced software engineers. Some data indicates that this could be one of the factors that’s recently been making it difficult for many of those younger professionals to find employment.
Proceed with caution
Despite their popularity, AI code-generation systems are far from perfect. They’re prone to all kinds of mistakes and unexpected behavior. In one illustrative debacle from July, Replit’s system deleted a user’s entire code database, and subsequently lied about it. Of that debacle, Replit CEO Amjad Masad, said on X that the company was addressing the specific causes of this issue to prevent a repeat, and that Replit would provide a beta feature to better insulate development environments.
Also: OpenAI’s fix for hallucinations is simpler than you think
The previous month, researchers from Anthropic found that many industry-leading AI agents will deceive and even threaten their human users if they perceive that their goals are being jeopardized.
Still, ZDNET’s David Gewirtz, an AI agent testing wizard, advised caution, noting the risks in letting an agent edit itself.
How to try it
You can try Agent 3 by signing up for an account with Replit here. The company offers a free trial for the new agent through its Starter tier. For full access, you’ll need to subscribe to the company’s Core service, which costs $20 per month. Teams and Enterprise tiers are also available.