Code is being written faster, but software isn't shipping any faster.

GitLab's latest developer survey, released this week, reached a conclusion embarrassing for many AI coding tool vendors: AI tools have significantly improved coding speed, but overall software delivery efficiency hasn't improved as a result. This bucket of cold water is well-timed — as Cursor's valuation is reportedly reaching $8 billion, Copilot becomes developer standard issue, and Claude Code floods social feeds, someone is finally asking the neglected question: then what?

The Bottleneck Shifted, It Didn't Disappear

GitLab's data reveals a classic “bottleneck shift” phenomenon. AI coding assistants excel at local tasks like code completion, function generation, and bug fixes; individual developer code output may be up 30-50%. But software engineering was never just about writing code — requirements analysis, system design, code review, integration testing, security scanning, deployment and operations… these are the stages that determine software delivery cycles.

Faster coding has instead shifted the bottleneck downstream. An AI-assisted developer who can write 500 lines a day might wait three days for a test environment; an auto-generated API endpoint might require three rounds of rework because it didn't understand business context. A contemporary InfoQ analysis put it bluntly in the headline: “Frontend and Backend Disappear Together: AI Coding is Rewriting Big Tech Engineer Roles”.

This also explains why Cloudflare CEO Matthew Prince warned: “Over the next two years, Agents will cause a Log4j-level security vulnerability on the internet every week.” The faster AI generates code, the greater the pressure on security review and quality assurance. When code output speed exceeds a team's review and validation capacity, technical debt and security vulnerabilities accumulate faster.

The Toolchain Fills In: From Writing Code to Shipping Code

The market is responding to this problem. This week Vercel launched Eve, an open-source AI Agent development framework, with the goal not of making AI write more code but of helping AI Agents complete full development-to-deployment workflows better. Dapr version 1.18 introduced “verifiable execution,” giving AI Agents and workflows cryptographic-grade trust capabilities — meaning you can verify that operations an AI claims to execute were actually executed, without tampering or injection.

Both tools point in the same direction: the second half of AI coding isn't about who writes code faster — it's about who can get AI to reliably complete end-to-end delivery tasks. From generating code snippets, to generating complete feature modules, to autonomously completing the full requirement-to-production flow — each step requires more reliable toolchain support.

For individual developers, this means skill stack priorities are shifting. Pure “code writing” ability is depreciating because AI already does it well; but system design ability, requirements understanding, cross-system integration, and quality control are appreciating. InfoQ's report noted that Big Tech engineer roles are being rewritten by AI, with traditional “frontend/backend/QA” boundaries blurring, replaced by roles closer to end-to-end business delivery.

The AI coding tool story is far from over. Cursor, Copilot, Cline, MarsCode have solved the “writing” problem, but in a complete software delivery lifecycle, “writing” accounts for less than 30% of time. The remaining 70% — testing, integration, deployment, ops, security — is the next battlefield. Whoever can AI-ify and automate those stages will be the real winner of AI coding.