But the competitive landscape for AI-assisted coding platforms is overcrowded. Startups Windsurf, Replit and Polishide also sell AI-Code-Generation instruments to developers. Cline is a popular Open Source alternative. GitHub’s Copilot, developed in collaboration with Openai, is described as a ‘pair programmer’ who has the auto-completion code and offers resources.
Most of these code editors rely on a combination of AI models built by large technical enterprises, including Openai, Google and Anthropic. For example, Cursor is built on top of Visual Studio Code, an open source editor of Microsoft, and Cursor users generate code by tapping into AI models such as Google Gemini, Deepseeek and Anthropic’s Claude Sonnet.
Several developers tell Wired that they are now running Anthropic’s coding assistant, Claude Code, with the cursor (or instead). Since May, Claude Code has offered different debugging options. It can analyze error messages, step-by-step problem solving, perform specific changes and conduct unit tests in code.
This can all raise the question: how buggy is AI-written code compared to the code written by fallible people? Earlier this week, the AI Code Generation Instrument Replit reportedly became scary and made changes to the user’s code, despite the fact that the project was freezing in a ‘code freezing’ or a break. Eventually, the user’s entire database deleted. Reprit’s founder and CEO said on X that the incident was “unacceptable and should never be possible.” And yet it was. This is an extreme case, but even small errors can cause a destruction for coders.
Anysphere does not require a clear answer to whether AI code requires more AI code -FAILING. Kaplan argues that it is “orthogonal that people encode a lot.” Even though all the code is written by a human, it is still very likely that there will be mistakes, he says.
Anysphere product engineer Rohan Varma estimates that as much as 30 to 40 percent of the code is generated by AI. This is in line with estimates shared by other companies; For example, Google said that about 30 percent of the company’s code is now proposed by AI and revised by human developers. Most organizations still make human engineers responsible for checking the code before being deployed. It is striking that one recent randomized control trial with 16 experienced coders suggested that it took 19 percent longer To complete tasks as when they could not use AI tools.
Bugbot is meant to make it worse. “The heads of AI with our larger clients are looking for the next step with the cursor,” says Varma. “The first step was:” Let’s increase the velocity of our teams, let everyone move faster. “Now that they are moving faster, it is:” How can we make sure we don’t introduce new problems, we don’t break things? ” ‘He also emphasized that bugbot is designed to detect specific types of errors logical errors, security issues and other cases.
One incident that bugbot ratified for the Anysphere team: A few months ago, the (human) codes on Anysphere realized that they had not received a comment from bugbot on their code for a few hours. Bugbot went down. Anysphere engineers began investigating the issue and finding the request to break the interruption for the interruption.
There in the logs, they saw that Bugbot commented on the request to warn a human engineer that if they had made this change, it would break the bugbot service. The instrument correctly predicted its own downfall. In the end, it was a person who broke it.