{"slug": "over-half-of-developers-now-write-less-than-20-of-their-code", "title": "Over Half of Developers Now Write Less Than 20% of Their Code", "summary": "JetBrains' 2026 Developer Ecosystem Survey of over 15,000 developers found that more than half of professional developers now manually write less than 20% of their code, with AI agents generating 46% of committed code on average and AI assistance contributing another 39%. The survey also revealed that 96% of developers do not fully trust AI-generated code, yet only 48% always verify it before committing, while Google's DORA data shows a 90% increase in AI adoption correlated with a 9% rise in bug rates and a 91% increase in code review time.", "body_md": "JetBrains surveyed more than 15,000 developers between May and July 2026. The headline: over half of professional developers now manually write less than 20% of their code. One in five writes zero lines without AI help. These are not forecasts. This is the current production baseline, and the numbers are moving in one direction.\n\n## The Numbers Are Already Extreme\n\nThe [JetBrains Developer Ecosystem Survey](https://blog.jetbrains.com/research/2026/08/how-much-code-do-developers-really-let-agents-write/) — the largest of its kind, in its tenth year — asked developers to report what percentage of their monthly code output was fully agent-generated, AI-assisted, or manually written. The results landed hard.\n\nOn average, 46% of committed code was fully generated by AI agents, another 39% was written with AI assistance, and just 27% was typed manually without help. Sonar’s independent [2026 State of Code survey](https://www.sonarsource.com/state-of-code-developer-survey-report.pdf) corroborates this: AI already accounts for 42% of all committed code across organizations, and Sonar projects that number will hit 65% by 2027.\n\n[Temporal’s 2026 State of Development Report](https://temporal.io/reports/state-of-development-2026) — surveying 554 engineers across the US and UK — found that 80.8% now use AI agents daily, up from 47.3% just twelve months ago. The median developer runs five agents. The average runs 10.7. Some teams run over 100.\n\n## More Code, Not Better Code\n\nHere is where the story gets uncomfortable. AI-generated code ships fast, but it arrives with baggage. Research puts AI-generated code at 1.7x more issues than human-written code. Forty-five percent of AI output contains OWASP Top 10 vulnerabilities — 2.74 times the rate of human-authored code.\n\nDevelopers know this. Sonar’s survey found that 96% of developers do not fully trust the code AI generates. The twist: only 48% always verify it before committing. The gap between knowing and doing is where incidents live.\n\nGoogle’s DORA data documented what that gap costs in production: a 90% increase in AI adoption correlated with bug rates climbing 9%, code review time ballooning 91%, and PR sizes growing 154%. The speed gains are real. So are the quality regressions. You cannot have the upside without owning the downside.\n\n## The Senior/Junior Paradox\n\nOne of the more surprising JetBrains findings: roughly a quarter of senior developers now generate 80% or more of their code via agents. Junior developers, by contrast, lean toward AI-assisted workflows but are less likely to be fully agentic.\n\nThis inverts the traditional assumption that automation is a crutch for the inexperienced. Experienced developers have the architectural judgment to direct agents effectively, specify tasks clearly, and catch errors during review. They automate more because they can do it well.\n\nThe concern is on the junior side. Developers who move to agent-first workflows before building manual debugging instincts may never develop the mental models that make recovery from agent failure possible. When the agent produces something subtly wrong — and 41.1% of developers report agent-related issues daily according to Temporal — understanding why requires grounding that shortcuts skip.\n\n## The Job Description Changed Without Announcement\n\n“The title software engineer is going to start to go away. It’s going to be replaced by ‘builder,’ and it’s going to be painful for a lot of people.”\n\nBoris Cherny, creator of Claude Code at Anthropic\n\nCherny [told Fortune](https://fortune.com/2026/02/24/will-claude-destroy-software-engineer-coding-jobs-creator-says-printing-press) he has not written a line of code himself in more than six months. Dario Amodei, Anthropic’s CEO, has made similar claims — that models doing all software engineering end-to-end may be six to twelve months away.\n\nThese predictions deserve scrutiny. The data shows agents can generate code at scale, but the skill of knowing what to build, how to structure it, how to secure it, and whether the agent’s output is actually correct — that judgment has not been automated. It has been made more important.\n\nThe skills that matter now: writing clear specifications, decomposing complex problems into bounded tasks, reviewing AI output with the same rigor you’d apply to a junior developer’s pull request, understanding agent failure modes, and being the one responsible when things go wrong. None of these are new skills. All of them are harder to develop if you only practice directing agents.\n\n## What to Actually Do\n\nThe practical response is not to write more code manually. It is to treat the review step as the job, not as overhead.\n\n**Treat agent output like a junior’s PR.** Read every line. Question patterns that look convenient but smell wrong. Agents optimize for plausibility, not correctness.**Write better specs.** Garbage in, garbage out applies more than ever. Invest in the specification, not the prompt.**Learn agent failure modes.** Understand how your stack behaves when agents stall, loop, or produce context-overflowed output. 41% of developers hit this daily — most teams have no runbook for it.**Run security scans on AI output.** With 45% of AI code containing OWASP Top 10 vulnerabilities, static analysis is not optional. It is the catch layer your workflow requires.\n\nThe risk is not that developers will be replaced this quarter. The risk is that teams ship faster with less understanding of what they are shipping — and discover the cost of that tradeoff during an incident at 2 AM.\n\nThe job description changed. The question is whether your instincts caught up.", "url": "https://wpnews.pro/news/over-half-of-developers-now-write-less-than-20-of-their-code", "canonical_source": "https://byteiota.com/over-half-of-developers-now-write-less-than-20-of-their-code/", "published_at": "2026-09-02 05:14:32+00:00", "updated_at": "2026-09-02 05:23:18.430923+00:00", "lang": "en", "topics": ["artificial-intelligence", "developer-tools", "ai-agents", "ai-products"], "entities": ["JetBrains", "Sonar", "Temporal", "Google DORA", "Anthropic", "Boris Cherny", "Dario Amodei", "Claude Code"], "alternates": {"html": "https://wpnews.pro/news/over-half-of-developers-now-write-less-than-20-of-their-code", "markdown": "https://wpnews.pro/news/over-half-of-developers-now-write-less-than-20-of-their-code.md", "text": "https://wpnews.pro/news/over-half-of-developers-now-write-less-than-20-of-their-code.txt", "jsonld": "https://wpnews.pro/news/over-half-of-developers-now-write-less-than-20-of-their-code.jsonld"}}