Salesforce researcher claims agentic coding eliminates manual code writing

Ben Stolovitz, a research engineer at Salesforce AI Research, published a post in October 2026 detailing his shift from manual coding to managing AI agents.reported

The claim

He describes a process where each agent receives a separate Git worktree to execute tasks independently. The agents write code, run tests, and create pull requests without direct line-by-line input from him. He claims this setup allows for rapid prototyping, with new features built in approximately twenty minutes.

The response

The post acknowledges that this approach is purely personal opinion and speculation. Stolovitz notes that he previously worked on Microsoft’s Windows AI team and now supports scientists at Salesforce. He admits that norms for using these tools are still forming and that humans are currently "messing up" as AI integrates into daily life. He expresses openness to his current methods being incorrect.

The practical upstop

The material leaves the broader industry impact unverified, focusing solely on one engineer’s experience. Stolovitz emphasizes that he still uses AI in similar ways to the previous year, but with increased automation. He highlights the need for a personal policy to navigate the blurring lines between human and machine output, citing examples of colleagues using AI for Slack messages.

Claims versus evidence

The source is a single individual’s blog post, not an independent audit or company report. Stolovitz explicitly labels his account as "purely my opinion & speculation." The claim that manual coding is "obsolete" applies only to his specific role as a research engineer building rapid prototypes, not to all software development. The $3,000 cost is self-reported and may not reflect average enterprise spending.reported No external validation of the twenty-minute prototype time is provided.