The AI coding agent hangover has begun
Companies raced to embrace Claude Code, Codex and other AI coding tools. Now they're confronting surprise bills, cognitive overload and a harder question: Is all that AI-generated code worth it?
For the past few weeks, I’ve been asking people deploying coding agents like Claude Code and Codex inside their companies: What happens when the honeymoon is over?
It turns out there’s a pretty big hangover.
Over the past six months, as the latest models have made coding agents increasingly capable, developers inside large companies have raced to put them to work.
But at companies that have already been using coding agents for several months, the mood is changing. The surprise token bills have been the most widely publicized parts of the hangover: For example, one IT manager at a mid-sized tech company recently wrote on Reddit that after rolling out AI coding tools to 500 developers, the company was on track to spend $340,000 a year. And it had no way to prove the ROI.
Jason Elrod, a healthcare CISO and executive advisor at cybersecurity company Elisity, described a version of this problem as a "denial of wallet attack": an AI agent that can consume resources at a speed and scale that simply isn't possible for a human employee.
Elisity learned this firsthand. CMO Charlie Treadwell recalled a surprise AWS Bedrock bill of about $30,000 early in the company's experimentation with Claude Code, when an engineer didn't realize how much he was consuming. "We didn't realize he did it until we got the bill at the end of the month," Treadwell said. The company immediately put spend controls in place and hasn't had a surprise bill since.
But there are plenty of other problems that pop up a few months after deployment. For example, companies are discovering security and governance gaps, and that their internal systems simply weren’t designed for autonomous software. Some observe developers confronting a new kind of cognitive overload as they supervise multiple agents at once. Other find that generating more code doesn’t necessarily mean producing more value.
From euphoria to reality
“Initially there is this euphoria of ‘wow, this is so cool,’ like a honeymoon period,” said Vlad Luzin, cofounder and CTO of Band, which is building a communications and governance system for companies’ AI agents. However, that is followed by a recognition of problems: “And the higher up you are the more time it takes for you to understand actually the implications,” he said.
According to one engineering VP at a major software platform in late 2025 he would have immediately said he was excited about code-generating new application features. “Fast forward to June 2026 and my own perspective of the models has changed,” he said.
One reason, he said, is the quality of the code itself. Claude Code and Codex can be remarkably confident that the code they generate is better than the five-year-old code they’re replacing, but “we found that that generated code requires a lot of work.”
Noe Ramos, VP of AI operations at software company Agiloft, described another problem: “Traditional software fails loudly,” she said, while “AI-generated code fails quietly.” The code changes may look plausible and compile successfully, she explained, only for the problem to surface two sprints later. “Without that foundation, automation just accelerates the mess.”
Tim Doll, CEO of tech consultancy Precocity, had a different way of describing this phenomenon: “We are living in the era right now of the prototype that everybody thinks is a product.”
Claude Code can quickly create something that looks functional, he said, making it easy to mistake a prototype for production-ready software. But the tool doesn’t supply the expertise required to turn one into the other.
“It doesn’t make you a software architect just because you have Claude Code,” Doll said.
When one developer becomes a manager of agents
Band’s Luzin pointed out that instead of working on one task at a time, developers may be supervising multiple agents simultaneously—one writing code, another reviewing it, and still others working on separate tasks in the background. That can create a kind of cognitive overload that wasn’t expected.
“It’s like driving a Formula One car. If you lose focus, you hit the wall,” he said, adding that the work can also become more siloed and harder for the organization to see. Managers may have little visibility into what agents are doing, he explained, what a particular task actually costs, or how employees are using them.
AI companies sell enterprises on increased productivity and creativity, Luzin said, and they are right. But what gets less attention, he argues, is “the lack of visibility... the lack of cost attribution, the impact on your workforce,” as well as the potential for greater siloing and cognitive overload.
