If software is eating the world, AI and agentic technology are making a push to revolutionize software development, as recent releases from the likes of Tabnine, Zencoder and Microsoft, among others indicate.
Coding agents are taking on tasks such as intelligent code generation, code repair, test generation, code reviews and real-time optimization.
For instance, Tabnine introduced its Tabnine Code Review Agent in preview as the first AI agent that can incorporate and enforce company-specific development standards, Peter Guagenti, president of Tabnine, told The New Stack.
The product converts plain language requirements into comprehensive review rules, reviews code both at the pull request stage and within the IDE and provides suggested fixes, not just flagging issues.
Guagenti said key differentiators of the Tabnine code review agent include personalization — as it adapts to each team’s methods and preferences, along with ease of use and comprehensive coverage, as it reviews against both company-specific and industry standard rules.
“Right now, these tools are very powerful, and they’re emerging really fast, but they’re behaving sort of like an engineer off the street,” Guagenti said. “Our mission at Tabnine is to have a product that behaves like an onboard engineer who knows your company, knows your team.”
When developers create a pull request, the Tabnine Code Review Agent checks the code in the pull request against the rules established by their team. If any aspect of the code doesn’t conform with those rules, then the agent flags it to the code reviewer, providing guidance on the issue and suggested edits to fix it, the company said.
“AI is already in use in limited ways reviewing and validating code. However, like the static code analysis tools that came before them, current AI tools have been limited to checking code against generic, predefined standards,” wrote Shantanu Kedar, senior director of product marketing at Tabnine, in a blog post. “The challenge is that every mature engineering organization has unique and intricate ways of creating software applications. What one team sees as their irrefutable standard, another team might reject outright.”
Guagenti said the code review agent offers automatic fixes rather than just identifying issues like static analysis tools. It supports over 600 languages and frameworks and uses various large-language models (LLMs) with custom prompt engineering.
The product is now in private preview with Tabnine’s enterprise customers, several with large enterprise engineering teams. Current customers include chip makers, military/government institutions, financial services, and pharmaceutical companies, he said, noting that the company will also offer a simplified version of the technology for individual developers via credit card purchase.
Tabnine uses a “three-legged stool” approach:
And the product includes a proprietary context engine that understands company-specific code patterns and standards
Guagenti said his vision is to provide developers with 10x productivity gains, whereas current tools achieve 20% to 50% gains. “They’re not even doubling yet,” he said.
Overall, the company positions itself as moving beyond generic AI assistance to provide contextually aware, organization-specific code review and development support, he said.
“A lot of time is lost in these engineering reviews. The most senior people are the ones who are doing these reviews,” Guagenti told The New Stack. “And even with those most senior people doing it, not only are they spending a ton of time on it, they’re still missing stuff.”
Meanwhile, Zencoder, a new AI coding assistant company, recently launched its platform and AI agents that compete with GitHub Copilot and other coding tools. The company focuses on production-ready code rather than demos and uses what they call “compound AI systems” that combine traditional development tools (compilers, linters, etc.) with AI capabilities, Andrew Filev, founder and CEO of the startup, told The New Stack.
After a year in development and 500 companies in early access, Zencoder has delivered features including code completion, repository-aware AI chat, and coding agents for tasks like unit testing.
The company predicts significant automation of routine engineering work within four years while emphasizing the continuing importance of human creativity and system thinking.
The Zencoder tools work with Visual Studio Code and JetBrains IDEs and support major languages like Java, C#, and JavaScript.
According to the company, Zencoder’s primary technology pillars include:
“We’re one year old, and we build the product that competes with the best… And it took them three years to build that product,” Filev told The New Stack, aiming his sights at GitHub.
Filev said he believes that in the next four years, we’ll get to the point where AI will be able to automate about half of all the routine work in engineering.
However, “It’s not just like, oh, you give it to LLM, and LLM gives you the perfect answer. That’s not how I see the industry evolving… I see the industry evolving… building what’s called compound AI systems.”