#24 - AI coding: Productivity paradox revealed

Speed vs. quality in AI-assisted development

Author

Ian L.
September 23, 2025

Hey readers! 👋 The AI coding tool landscape is evolving rapidly, but there's a fascinating productivity paradox emerging. While AI generates code faster than ever, we're seeing bottlenecks in review processes and quality control. This week, we explore the tension between speed and quality in AI-assisted development, plus the explosion of CLI tools competing for your terminal. Let's dive in! (Ed: "workslop" is my new favorite term for low-quality AI output!)

This Week's Highlights 🔍

The Productivity Paradox of AI Coding

AI-Generated "Workslop" Is Destroying Productivity — Despite widespread adoption of generative AI tools in the workplace, companies are experiencing a paradox where increased AI usage does not translate into measurable productivity gains. – @harvardbiz

A confusing contradiction is unfolding as companies embrace AI coding tools: while AI-led processes have nearly doubled last year, an MIT Media Lab report found that 95% of organizations see no measurable return on their AI investments. The challenge isn't generating code—it's ensuring that code creates real value rather than producing what's now termed "workslop."

The productivity paradox of AI-assisted coding — AI-assisted coding tools have significantly accelerated code generation but created bottlenecks in code review, integration, and testing. – Edgar Kussberg

A METR study revealed that AI tools actually decreased experienced developers' productivity by 19%, despite developers perceiving time savings of about 20%. The surge in AI-generated code volume strains human reviewers, with 100% of developers reporting they need to modify AI-generated code. To truly benefit from AI, engineering teams must evolve their workflows toward sustainable, high-quality throughput rather than raw speed.

Vibe coding has turned senior devs into 'AI babysitters,' but they say it's worth it — Senior developers find themselves acting as "AI babysitters," spending significant time fixing AI-generated errors. – Dominic-Madori Davis

"Vibe coding is akin to 'hiring your stubborn, insolent teenager to help you do something.' You have to ask them 15 times to do something... they break a bunch of things along the way," says Feridoon Malekzadeh. Despite these drawbacks, many developers acknowledge that AI coding tools increase productivity and are becoming an integral part of software development workflows, with the caveat that human oversight remains essential.

The Rise of CLI-Based AI Coding Tools

CodeRabbit CLI | AI Code Reviews in CLI — CodeRabbit CLI is an AI-powered code review tool that integrates directly into the terminal, enabling developers to perform senior-engineer level code reviews before commits and pull requests. – CodeRabbit

This new CLI tool provides line-by-line reviews, one-click fixes, and context-aware feedback that helps catch bugs, logical errors, and code smells early in the development workflow. It integrates with AI coding assistants like Claude Code, Cursor CLI, and Gemini for seamless generate-review-iterate cycles. (Ed: CLI tools are everywhere. Is the Cursor IDE in trouble?)

Gemini CLI Unlocks New Developer Productivity Horizons — The Gemini CLI leverages a massive 1 million token context window to analyze entire project directories, enabling rapid comprehension of project architecture and dependencies. – StartupHub.ai Staff

Developed by Taylor Mullen, this groundbreaking AI-powered command-line interface addresses common challenges such as onboarding onto unfamiliar codebases. Beyond onboarding, the tool automates tasks like code health checks, refactoring suggestions, and Git history analysis, with extensibility through Model Context Protocol (MCP) servers allowing integration with external services.

OpenAI Releases GPT-5-Codex Optimized for Complex Code Refactoring and Code Reviews — OpenAI has launched GPT-5-Codex, a specialized version of GPT-5 designed for complex software engineering tasks such as large-scale code refactoring and extended code review workflows. – Contacted

GPT-5-Codex features adaptive reasoning, allowing it to allocate more computational effort for complex tasks while operating autonomously for extended periods. For the simplest 10% of requests, it used 93.7% fewer tokens than GPT-5, while for complex tasks, it spent roughly twice as long on reasoning, editing, testing, and iteration. The model achieved 51.3% accuracy on multi-step refactoring tasks versus 33.9% for GPT-5.

Learning and Workflow Integration

The beginner's guide to coding with Cursor — Lee Robinson, Head of AI Education at Cursor, explains how Cursor's AI-powered code editor helps both beginners and experienced developers. – Claire Vo

Cursor's AI agent can automatically detect and fix linting errors without requiring complex terminal commands. Setting up typed languages, linters, formatters, and tests creates guardrails that improve AI-generated code quality. A custom prompt with banned words and phrases can improve writing by eliminating common AI-generated patterns and clichés. (Ed: all these tools have learning curves. Send this to your favorite dev that's new at cursor)

Why Linear Built an API For Agents — Linear built a specialized API to integrate AI coding agents like Cursor directly into its platform. – Frederic Lardinois

This integration allows developers to assign tasks to the Cursor agent, which can then work autonomously in the background to fix issues or implement changes. Linear's API includes an 'agent session' concept that provides context and streamlines interactions between the agent and the platform. "It's not unusual, even on our own team, to see an issue that has just a title, and then someone says: '@cursor, fix this.' What did you expect to happen exactly here? Like these things aren't magic."

AI Code Review Tools 🔍

ChatGPT Code Reviews: How AI Feedback in 3 Minutes Beat Human Reviews Every Time — ChatGPT caught 312 bugs missed by senior developers in 847 pull requests. – @naikpratham230

AI code reviews reduce development cycle times from days to minutes by providing instant feedback. ChatGPT applies consistent standards regardless of time or reviewer fatigue, eliminating variability in human reviews. The hybrid approach—using AI for initial reviews and humans for strategic decisions—significantly reduces review time, with real teams seeing deployment frequency increase by 300% and production incidents drop by 25%.

On the Use of Agentic Coding: An Empirical Study of Pull Requests on GitHub — This empirical study analyzed 567 pull requests generated by Claude Code across 157 diverse open-source GitHub projects. – project maintainers

The results indicate that 83.8% of agent-assisted PRs are eventually accepted and merged by project maintainers, with 54.9% integrated without further modification. However, 45.1% of merged PRs require additional human revisions, especially for bug fixes, documentation, and project-specific standards. These findings suggest that while agent-assisted PRs are largely acceptable, they still benefit from human oversight and refinement.

Anthropic's NEW Claude Code Review Agent (Full Open Source Workflow) — Anthropic's Claude Code team has replaced manual line-by-line code reviews with AI agents, significantly speeding up the review process. – Patrick Ellis

"Anthropic just killed the manual code review. The engineers building Claude Code, which itself is nearly 95% written by Claude Code, no longer review most of their changes line by line." The review process is split into an AI-driven automated review (inner loop) and a human strategic review (outer loop) focusing on high-level architecture and business goals. Anthropic has open-sourced repositories including GitHub Action runners and security review tools.

Tool Comparisons and Announcements 🛠️

Which AI Coding Assistant Dominates in 2025: Codex vs Claude Code vs Cursor vs Copilot — GitHub Copilot integrates tightly with VS Code, supporting 50+ languages and improving coding speed by up to 70% for boilerplate tasks. – Ashley

Claude Code operates as a terminal-native AI assistant, excelling in multi-file reasoning and complex project tasks. Cursor offers an AI-native development environment with conversational AI and advanced collaboration features but requires significant workflow migration. All platforms achieve high syntax accuracy but vary in semantic correctness and best practice adherence; human review remains essential for security.

Visual Studio 2026 doubles down on AI-assisted coding — Microsoft has released the first preview of Visual Studio 2026 featuring deeper integration with GitHub Copilot AI. – InfoWorld

The update includes a Profiler Agent that helps developers analyze and optimize performance without expert knowledge, improved code search, and the ability to bring custom language models to Visual Studio Chat. The IDE also boasts a modern, lighter design aligned with Fluent UI, new color themes for accessibility, and an adaptive paste feature that adjusts pasted code contextually.

Cline for JetBrains is Here — Cline, a model-agnostic and inference-agnostic AI coding assistant, has expanded its platform support to include JetBrains IDEs. – @nickbaumann_

"Every major JetBrains IDE – IntelliJ IDEA, PyCharm, WebStorm, Android Studio, GoLand, PhpStorm, and more – now runs Cline natively." This update allows developers who prefer JetBrains environments to use Cline without switching IDEs, with full functionality including Plan and Act modes and seamless use of JetBrains' refactoring tools.

Quick Bytes 📱

Made with ❤️ by Data Drift Press

Have questions, comments, or feedback? Just hit reply—we'd love to hear from you!