Artificial Intelligence

Agentic Coders Are Rapidly Displacing Humans

The software development lifecycle will soon be dominated by autonomous agents, experts predict, but security and technical debt risks loom for organizations that move too fast.

Tara E. Buck

Tara worked as a newspaper reporter and editor for more than 13 years and served as executive editor of the full suite of CDW Tech magazines until 2019. She enjoys learning about new technology and traveling with her family.

Agentic coding workflows will soon transform software development from an environment that’s AI-assisted to one where agents dominate the entire lifecycle, from planning to creating and reviewing code, experts say. To cite one prediction, Gartner expects that more than 65% of engineering teams will use agentic coding by 2027, treating “integrated development environments as optional, and shifting control, governance and validation to automated platforms.”

“Agents continue to get better, and in sort of a dramatic way with respect to coding,” says Arnal Dayaratna, research vice president for software development at IDC. Today, agents are “capable not only of coding but also increasingly testing deployment at continuous integration. Pretty much at every step of the lifecycle, we’re seeing a lot of capabilities advance.”

Still, there’s room for improvement, he adds, and human developers must also expand their skill sets to manage and fully optimize the output their coding agents deliver.

“It’s an exciting time,” Dayaratna says. “It’s also immensely disruptive.”

Click the banner below to learn more about how your business can turn data into insights.

How Will Agentic AI Affect Human Coders?

Organizations that laid off software engineers have started to rehire them “because they overestimated what the AI productivity gains were going to be,” Dayaratna says. Technical debt stands as a looming risk for enterprises that may be unable to maintain early agentic outputs because they don’t understand how they were built. “It’s a particular risk simply because they may not staff to sift through the volume of code that’s generated or manage it and curate it in quite the same way.”

The sheer volume of code that agents can deliver also brings risk when it’s not tested in the same ways as human-generated code has historically been tested. “We haven’t set up infrastructures yet that enable agents to reliably maintain code and service it,” Dayaratna says.

Today, developers overwhelmingly rely on off-the-shelf agents, but Dayaratna sees a place for open models in the ecosystem that are hosted on-premises. Open-model examples include NVIDIA Nemotron and Mistral.

“The tools will just become better. Who’s going to win? Developers and enterprises. We’ll get to a point where the tools will become incredibly powerful, and we’ll do things faster and more reliably than we’ve ever done.”

WATCH: How NVIDIA is helping enterprises realize ROI and scalability goals for AI projects.

Application modernization and transformation is one area that’s ripe for improvement, Dayaratna says. “Modernizing legacy application estates, transitioning code from one language to another or upgrading software to a new language and framework — we frankly haven’t been able to do that very well. With agents, we’re entering a new world of development and software engineering.”

But risks abound, particularly the need to harden guardrails and security, which has already proved more difficult than some observers anticipated. In July, two OpenAI models escaped from a controlled test environment sandbox and autonomously hacked into another AI company, Hugging Face, to cheat on an evaluation test. After the models gained internet access, they hacked into Hugging Face servers by exploiting exposed credentials and zero-day vulnerabilities. The company had identified and contained the attack before OpenAI detected and alerted teams of the exploit, according to reporting from Fortune magazine. Both companies are investigating the issue.

elenabs/Getty Images