AI coding addiction burnout has moved from anecdote to data point: a survey of 305 developers published by Coddy Tech on 13 August 2026 found that four in five (80%) said their use of AI coding tools had felt more like dependence than an advantage.
When the Agent Loop Becomes a Treadmill
Quentin Rousseau, CTO and co-founder of incident-response company Rootly, captured the dynamic in a blog post that circulated widely among developers. ‘It’s 2:47 a.m… I’m not debugging an outage. There’s no deadline. I’m just watching Claude Code refactor a module… and I can’t stop,’ he wrote. Rousseau’s explanation for the pull: ‘Agentic coding is addictive. When the agent gets things right, you get a dopamine hit. When it fails, you get an adrenaline rush.’
He later told Axios that he could not sleep for months after switching to agentic coding and that a doctor eventually prescribed sleep medication to shut his brain off. In a post on his own blog, Rousseau compared agentic tools to slot machines: ‘You hit one prompt, you get an answer, you get some coding done.’
The Coddy Tech survey, conducted in July 2026 among developers who use AI coding tools at least once a week, puts numbers to that experience. According to the Coddy Tech AI Coding Addiction Report, 43% of developers keep coding with AI after hours even when they intended to stop, and 32% have put off sleep to keep going. A further 39% said AI tools had made it harder to switch off from work altogether.
Among specific tools, Startup Fortune’s summary of the Coddy data shows Claude Code was named by 35% of surveyed developers as the hardest tool to put down. Codex users stood out on after-hours work: 62% of them said they kept going past the point they had meant to quit.
AI Coding Addiction Burnout and the Verification Debt Problem
The survey found that 74% of developers believed heavy AI use made them more likely to earn a raise or promotion. But 51% said the same heavy use made them more likely to burn out, a figure consistent with the report’s finding that a sub-group it labels ‘Hooked’ developers was far more likely than others to say work had become hard to disconnect from.
The productivity case for AI tools does not fully account for a problem the industry is calling verification debt. Code arrives quickly, but developers still need to judge whether it is correct, secure, maintainable, and appropriate to the specific codebase. The 2025 Stack Overflow Developer Survey, which drew more than 49,000 responses from 177 countries, found that 66% of developers are spending more time fixing ‘almost-right’ AI-generated code, and that 75% said they would still turn to another person for help when they did not trust the AI’s answer.
Trust figures are deteriorating. The Stack Overflow survey’s AI section found that more developers actively distrust AI accuracy (46%) than trust it (33%), with only 3% reporting they ‘highly trust’ the output. Experienced developers, those with ten or more years of professional experience, showed the lowest ‘highly trust’ rate at 2.6% and the highest ‘highly distrust’ rate at 20%. The Stack Overflow blog reported that trust in AI accuracy has dropped from 40% in 2023 to 29% in 2025, even as adoption climbed from 76% in 2024 to 84% in 2025.
The Stack Overflow press release for the survey also noted that positive sentiment toward AI tools has fallen from 72% to 60% year on year. Developers are using the tools more, trusting them less, and feeling worse about the overall experience.
There is a structural reason for that gap. If an organisation treats AI as a multiplier of developer capacity rather than a reducer of toil, workers face pressure to ship more features, close more tickets, and review more pull requests within the same number of hours. That pressure erases the time saved on individual coding tasks and relocates it: larger pull requests, more generated changes to inspect, more dependencies to validate.
Rousseau’s situation illustrates where that leads when the work expands into time that used to belong to sleep. The question facing teams is not whether AI can write code. The practical reckoning is whether organisations will allow developers to draw a line around when the agent loop stops.