Software developers are no longer spending most of their week writing code. New survey data released this month shows the opposite: nearly 80% of developers now spend less than half their work week producing new code from scratch, while the share who say artificial intelligence writes at least half of their code has more than tripled in a single year, climbing from 12% to 42%. The findings, drawn from nearshore development firm BairesDev’s Q3 2026 Dev Barometer survey, mark one of the clearest signals yet that the job of “software engineer” is being redefined around reviewing, testing, and approving AI output rather than typing it line by line.
The shift is not a minor adjustment. It touches hiring, tooling budgets, and the basic skills companies now look for when they bring in engineers. It also arrives alongside separate data from Stack Overflow, reported by InfoWorld in October 2026, showing that developers still don’t fully trust AI output even as they rely on it more than ever. Taken together, the numbers describe an industry moving from code production toward code supervision, with real consequences for how software gets built, who ships it, and what tools companies like GitHub, Anthropic, and OpenAI are racing to sell into that gap.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
The Numbers Behind the Shift: 12% to 42% in a Year
BairesDev fielded its Q3 2026 Dev Barometer in August 2026, surveying 705 software developers across more than 60 countries and, for the first time, 41 enterprise chief technology officers from Fortune 500 and mid-market organizations. The headline number: 42% of developers now say AI writes at least half of their code, up from just 12% in the same quarter a year earlier, according to BairesDev’s published executive summary and corroborating coverage from Yahoo Finance and VentureBeat.
That is not a gradual climb. It is a more than threefold jump in twelve months, and it lines up with the broader rollout of coding-focused AI tools over the same window, including Claude Code, Codex, and GitHub Copilot all expanding their agentic capabilities through 2026. VentureBeat’s Carl Franzen, who separately reported on the same survey, framed the jump as evidence that AI-generated code has moved from a novelty feature to the default starting point for a large share of professional engineers.
The distribution of answers matters as much as the headline figure. BairesDev’s data shows the largest single group of developers, 30% of respondents, say AI contributes somewhere between one-quarter and one-half of their code. That is the modal answer, not an outlier. It suggests most developers are operating in a blended mode, where AI handles a meaningful chunk of output but human judgment still shapes a large share of the final product.
Developers Are Writing Less, Not Doing Less
The flip side of the 42% figure is what developers are doing with the time AI frees up. Only 21% of developers surveyed said they spend more than half their work week writing new code from scratch, according to Yahoo Finance’s coverage of the Dev Barometer data. That means close to 80% of developers now spend the majority of their week on something other than original code authorship.
BairesDev characterized that reallocated time as going toward review, validation, testing, and decisions about whether AI-generated code is safe to ship. Crucially, the survey data indicates the final call on what goes to production still rests with a human engineer, not the AI tool itself. That distinction is the center of the current debate inside engineering organizations: AI can draft code at scale, but accountability for what ships has not moved.
There is also a wellbeing angle buried in the numbers that has gotten less attention than the adoption statistics. BairesDev’s year-over-year comparison shows developer-reported role fulfillment rose from 76% to 86% over the same twelve-month period covered by the survey. If that holds up in future survey waves, it would undercut the narrative that AI-driven code review work is purely a downgrade from “real” engineering, suggesting instead that many developers see the shift toward oversight and architecture as a net positive for job satisfaction, not a demotion.
Why CTOs Are Watching This Closely
The inclusion of 41 enterprise CTOs in this wave of the Dev Barometer is itself notable. BairesDev’s prior surveys tracked developer sentiment directly; adding a management-level cohort lets the firm compare what individual engineers report with what their bosses are planning to fund. Enterprise leaders do not need developers to feel good about AI-written code. They need predictable delivery, lower defect rates, and tooling that scales review capacity as fast as AI scales output.
That is precisely the gap that companies such as GitHub are now building products to close. GitHub has introduced a public-preview capability that lets its Copilot code review feature approve pull requests and satisfy required-approval rules automatically, according to an October 2026 InfoWorld report. GitHub has paired that capability with an explicit warning: automated reviews can miss problems and should be supplemented by human review, not replace it outright. The company is effectively building toward a world where AI both writes the code and performs the first pass of review, with humans positioned as a secondary check rather than the primary gate.
That workflow, where an issue or a prompt becomes a proposed change, then a pull request, then an item for further human review, is becoming the default pattern across GitHub’s ecosystem of first-party and third-party agents, including tools built on Anthropic’s Claude and OpenAI’s Codex models. Developers are increasingly supervising agent-produced changes in bulk rather than authoring every line individually, a workflow pattern that mirrors what teams building on cloud-native AI gateways have already had to design around for production traffic.
Trust Hasn’t Caught Up to Adoption
Here is where the story gets more complicated than a simple adoption curve. Stack Overflow’s trust data, as reported by InfoWorld in October 2026, shows developers are using AI tools constantly while still not fully trusting them. Only 7% of developers say they trust AI for important work decisions without being able to verify the output themselves. That figure rises to 48% when developers can easily check the AI’s work before relying on it.
InfoWorld notes that Stack Overflow’s 2025 survey asked a differently worded question, where roughly 33% of developers said they trusted AI output and 46% said they distrusted it. Because the question wording changed between the two survey years, InfoWorld cautions against treating the 2025 and 2026 figures as a clean year-over-year trend. What is consistent across both years, though, is the underlying pattern: trust in AI-generated code is conditional on the ability to verify it, not unconditional acceptance. That is the exact behavior BairesDev’s data describes developers exhibiting in practice, spending less time writing and more time checking.
This verification-dependent trust model has direct implications for how teams are restructuring their workflows, including those building agentic coding setups with tools like Cursor Composer’s diff-versus-rewrite modes, which exist specifically to make AI-generated changes easier for a human to audit before they merge.
Historical Context: From Autocomplete to Autonomous Agents
The move from writing to reviewing did not happen overnight. GitHub Copilot launched as an autocomplete-style suggestion tool, offering single lines or short blocks of code that a developer would accept, reject, or edit one suggestion at a time. That model kept the developer in the driver’s seat for every keystroke. Over the past two years, the dominant coding tools shifted toward agentic workflows: a developer describes a task, and the tool generates, tests, and sometimes commits entire features or bug fixes with minimal line-by-line supervision.
That architectural shift, from suggestion to delegation, is what made the jump from 12% to 42% possible in just one year. It also explains why tooling categories built around review and oversight, rather than generation, have expanded fastest. Editors and IDEs have adapted in parallel. VS Code’s Copilot agent sandbox and competing tools from Zed and Cursor have all added features aimed specifically at making it faster to audit what an agent changed, rather than faster to generate more code in the first place.
Competitive Landscape: Who Benefits From the Review Shift
The competitive stakes of this shift are significant because the tool that owns the review step effectively owns the final gate before code ships. That is a more durable position than owning the generation step, since generation is increasingly commoditized across multiple competing models. The table below compares how the major coding-tool players are positioning around the review-versus-generation divide as of October 2026.
| Tool / Platform | Primary Strength | Review-Focused Feature | Positioning in the Shift |
|---|---|---|---|
| GitHub Copilot | Deep GitHub/PR integration | Automated PR review and approval (public preview) | Owns both generation and the approval gate |
| Claude Code (Anthropic) | Long-context reasoning on large codebases | Agentic planning with human checkpoints | Leans on trust via transparency of reasoning |
| Codex (OpenAI) | CLI-driven autonomous task execution | Sandboxed execution before merge | Competes on task completion speed |
| Cursor | Editor-native agent workflows | Composer diff view vs. full rewrite mode | Built around making AI changes auditable |
| Zed | Lightweight, fast multi-model editor | Bring-your-own-key model switching | Lets teams choose the model they trust most |
GitHub’s approach, letting Copilot both write and approve code under certain conditions, is the most aggressive bet on automating the review layer itself. It is also the most contested, given GitHub’s own warning that automated review can miss problems. Anthropic and OpenAI, by contrast, have focused more on making the reasoning behind agentic changes legible to a human reviewer, betting that transparency rather than automation is what will earn developer trust at the review stage. The TIOBE Index’s October 2026 language rankings, which still show Python and established languages dominating despite the AI-authorship shift, suggest the underlying languages developers review have not changed even as who writes the first draft has.
Market Impact: Where the Money Is Moving
BairesDev’s data shows enterprise CTOs are not waiting to see how the trend plays out before adjusting budgets. The firm’s inclusion of CTO respondents for the first time in this survey wave points to a parallel management-side response: organizations are reallocating spend toward the infrastructure needed to validate AI output at scale, rather than purely toward tools that generate more of it. That tracks with GitHub’s own product direction, building automated review and approval features rather than only expanding code-generation capacity.
Vendors selling pure code-generation capability are increasingly competing on a commoditized axis, speed and context window size, while vendors building the review and verification layer are carving out a less contested, higher-margin niche. That dynamic mirrors what happened in other software categories where automation outpaced the trust needed to deploy it unsupervised, including the current build-out around GitHub Copilot CLI’s computer-use features, which several rival agent platforms have moved quickly to match.
The Skills Employers Are Now Screening For
If developers are spending less time writing code from scratch, the skills that make someone valuable on a team are shifting too. Code review literacy, the ability to quickly spot subtle logic errors or security issues in a large AI-generated diff, is becoming as important as the ability to write clean code unassisted. So is the judgment to decide when AI-generated code is not worth the time it would take to properly verify it, and should simply be written by hand instead.
That is a different skill set than what most computer science curricula have historically emphasized, and it is part of why engineering hiring conversations in 2026 increasingly center on how candidates work alongside AI tools rather than purely on algorithmic problem-solving. Competing CLI-based tools, compared directly in coverage of Codex CLI versus Claude Code, illustrate how much of the current tooling race is now about workflow ergonomics for review and iteration, not just raw code-generation quality.
What BairesDev’s Data Doesn’t Tell Us
It is worth being precise about what this survey does and does not establish. BairesDev’s Dev Barometer is a self-reported survey of 705 developers and 41 CTOs, not a code-analysis study measuring actual lines of AI-generated code merged into production repositories. The 12%-to-42% figure reflects what developers believe about their own AI usage, which is a meaningful signal of sentiment and workflow change but not a precise technical measurement.
Similarly, the available reporting does not yet include verified 2025-2026 figures from other major industry trackers, such as GitHub’s Octoverse report, independent AI-capability evaluators, or academic studies on AI-code defect rates, that would let the industry quantify exactly how much review time increases per AI-generated line, or how AI-introduced bugs compare to human-introduced ones at scale. Until that data arrives, the BairesDev and Stack Overflow figures should be read as directional evidence of a real shift, not as a complete measurement of its size.
Five Predictions for the Next Twelve Months
- The 42% figure will keep climbing in BairesDev’s Q3 2027 Dev Barometer, likely surpassing 50% as agentic tools handle larger, multi-file tasks rather than isolated functions.
- More platforms will follow GitHub’s lead in offering automated PR approval, but enterprise adoption of fully automated merges will stay limited due to liability and compliance concerns.
- Stack Overflow’s trust gap between verifiable and unverifiable AI output will persist, keeping demand high for tools that make AI reasoning auditable rather than opaque.
- Job postings for mid-level engineering roles will increasingly list code-review and AI-output validation as explicit required skills, not implied ones.
- A dedicated, independently audited benchmark for AI-code defect and security-vulnerability rates will emerge as the missing data point the industry needs to move past self-reported survey figures.
What This Means for Individual Developers
For developers reading their own week through this data, the practical takeaway is less about panic and more about recalibration. The BairesDev survey’s finding that role fulfillment rose from 76% to 86% over the same period suggests the shift, so far, is not translating into widespread dissatisfaction. Developers appear to be finding that spending more time on review, architecture, and validation, rather than raw code production, is not inherently a worse use of their expertise.
That said, the Stack Overflow trust figures are a useful check on overconfidence. With only 7% of developers trusting AI output on important decisions without verification, the clear expectation across the industry remains that a human stays accountable for what ships. Tools that make that verification faster and more reliable, rather than tools that try to remove the human step entirely, are where the current competitive energy is concentrated.
Data Snapshot: The Dev Barometer Numbers at a Glance
| Metric | Q3 2025 | Q3 2026 | Source |
|---|---|---|---|
| Developers saying AI writes 50%+ of their code | 12% | 42% | BairesDev Dev Barometer |
| Developers spending 50%+ of week writing new code | Not directly comparable | 21% | BairesDev Dev Barometer / Yahoo Finance |
| Developer-reported role fulfillment | 76% | 86% | BairesDev Dev Barometer |
| Developers trusting AI output only when easily verifiable | Question reworded | 48% | Stack Overflow via InfoWorld |
| Developers trusting AI for important decisions unconditionally | Question reworded | 7% | Stack Overflow via InfoWorld |
Frequently Asked Questions
What is the BairesDev Dev Barometer?
It is a recurring survey from nearshore software development firm BairesDev that tracks how developers are using AI coding tools over time. The Q3 2026 edition surveyed 705 software developers across more than 60 countries and, for the first time, 41 enterprise CTOs.
Does this mean AI is writing 42% of all code in production today?
No. The 42% figure reflects the share of surveyed developers who personally say AI writes at least half of their own code, based on self-reported survey responses. It is a sentiment and workflow indicator, not a measured count of AI-generated lines across the industry.
Are developers less busy now that AI writes more code?
The data suggests developers are busy differently, not less busy. Nearly 80% of developers spend less than half their week writing new code from scratch, with that time shifting toward review, testing, and validation of AI-generated output.
Do developers trust AI-generated code?
Only partially. Stack Overflow data reported by InfoWorld shows just 7% of developers trust AI output for important decisions without being able to verify it, while 48% trust it when verification is easy. Trust is conditional on the ability to check the work, not automatic.
What is GitHub doing about AI code review?
GitHub has introduced a public-preview feature letting Copilot code review approve pull requests and satisfy required-approval rules automatically, while cautioning that automated reviews can still miss problems and should be supplemented by human oversight.
Is this trend unique to any one AI coding tool?
No. The shift toward review-heavy workflows spans the ecosystem, including GitHub Copilot, Anthropic’s Claude-based coding tools, OpenAI’s Codex, and editor-native agents in Cursor and Zed. The common thread is tooling built around auditing AI output, not just generating more of it.
Why did developer role fulfillment increase if they’re writing less code?
BairesDev’s survey data shows fulfillment rose from 76% to 86% over the same period the code-writing shift occurred, suggesting many developers see a move toward review, architecture, and higher-judgment work as a positive change rather than a loss of core engineering responsibility.
What should developers do to stay competitive given this shift?
Building strong code-review skills, including spotting subtle logic errors and security issues in large AI-generated diffs, is becoming as valuable as writing code unassisted. Familiarity with the review-focused features of tools like GitHub Copilot, Claude Code, Codex, and Cursor is increasingly part of baseline engineering competency.
Related Coverage
- VS Code 1.140 Adds HydraFusion AI Routing, 4 Tiers [2026]
- AI Coding Agents Add 30% More Code, Not Software [2026]
- Build a Cursor Hooks VS Code Extension: 13 Steps [2026]
- Build a VS Code Ghost-Text AI Extension: 13 Steps [2026]
- Cursor vs Devin Desktop vs VS Code Copilot: $190 Plan Gap [2026]
