Augmented Coding Weekly

Claude Opus 4.5 achieves breakthrough 80% pass rate on SWE-Bench as foundation models continue rapid improvement despite scaling law concerns, while practical experiments reveal Spec-Driven Development's limitations compared to lightweight iterative approaches. A shockingly simple prompt injection attack on Google's Antigravity highlights the urgent security challenges as agentic AI tools gain deeper access to codebases and credentials.

Google releases Gemini 3.0 Pro and their own agentic IDE Antigravity, while the industry debates specification-driven development approaches. Plus, running AI agents in continuous loops and Martin Fowler's perspectives on AI's impact on software engineering.

What happens when code becomes cheap and abundant? This week explores the shift from code scarcity to wisdom scarcity, OpenAI's GPT-5.1 incremental improvements, and ByteDance's $1.30 coding agent that highlights how heavily subsidised our AI future really is.

AI can complete just 2.5% of real-world tasks autonomously, yet its adoption is already reshaping hiring patterns, with companies reducing junior roles while increasing demand for ML and data engineering positions.

This issue highlights Claude Skills for Neo4j, a reminder that AI can code but not build software, Google's new vibe-coding feature for rapid prototyping in AI Studio, and Cursor's Composer model release, which offers little transparency.

This issue explores Anthropic's new Agent Skills feature and its implications for MCP, examines the programmer identity crisis in the age of AI, reviews a developer's experience building a production app with Devin, and compares community sentiment between Claude Code and Codex.

In this issue we explore how AI is democratizing software development by enabling anyone to build "home-cooked" applications, examine the risks of "vibe coding hell," track AI tool adoption trends, analyze Augment Code's pricing crisis, and share one developer's optimistic journey with AI-augmented development.

This issue explores Claude's new Imagine prompt-to-prototype tool, Simon Willison's thoughts on parallel coding agents and "vibe engineering" practices, and discusses two key limitations of current LLM coding agents - their inability to cut/paste code and reluctance to ask clarifying questions.

Anthropic releases Claude Sonnet 4.5, positioning it as "the best coding model" with significant improvements in coding, reasoning, and computer use capabilities. Features include an upgraded Claude Code with Agent SDK, revamped terminal, and new VS Code extension.

Exploring advanced context engineering for AI coding agents in complex codebases, plus CompileBench - a new benchmark challenging AI models to revive 22-year-old code and tackle cross-compilation tasks that test genuine problem-solving abilities.

How to turn Claude Code into domain-specific coding agents using MCP servers and configuration files, plus insights on subagents for parallel development and the latest upgrades to OpenAI's Codex model.

Exploring the framework wars in agentic development, GitHub's new spec-driven development toolkit for AI, a three-month Claude experiment that created the "Cursed" GenZ programming language, and why reading code remains the real bottleneck in LLM-powered development.

Introduction of Agent Client Protocol (ACP) by Zed for plugging coding agents into editors, Claude Code integration into Zed, a veteran developer's in-depth exploration of vibe coding benefits and challenges, and the mixed results of top-down AI tool mandates in organizations.

A developer's rationale for declining AI-generated merge requests and the learning value of code review, a free workshop on building coding agents with surprisingly simple architecture, enterprise challenges with vibe debugging and the need for AI-proof development pipelines, and a GitHub project requiring AI tool disclosure for contributions.

Exploring how AI tools enable "doing things that don't scale" by keeping projects small and sustainable, a critical look at OpenAI's AGENTS.md format and its limitations compared to human-readable documentation, and a humorous video about the realities of vibe coding from a senior engineer's perspective.

Analysis of specialized app generation platforms and their security challenges for non-technical users, a deep dive podcast with Replit's CEO on vibe coding's potential to democratize software creation, a technical author's identity crisis in the age of AI explanations, and exploring Claude Code as a comprehensive development solution.

GPT-5 launch with focus on coding improvements and SWE Bench performance gains, exploring why typed languages excel in AI-assisted development, a malicious AI prompt injection attack in AWS Toolkit VSCode, and debunking the myth that AI makes engineers 10x more productive by examining real-world software development complexities.

A Ruby developer's experience using AI to break programming language barriers and learn C and Rust, Stanford research measuring real AI productivity gains in enterprise development teams, Anthropic's internal case studies on Claude Code usage, and exploring local AI models like GLM-4.5 for coding on modest hardware.

Antirez shares practical insights on coding with LLMs and prefers direct model interaction over integrated tools, exploring how to redesign CLI interfaces for better AI agent usability, the controversial Replit production database deletion incident at SaaStr, and a human programmer beating AI at the World Coding Championship.

Cognition's acquisition of Windsurf after a complex series of deals with OpenAI and Google, survey results showing GitHub Copilot's enterprise dominance and the rise of Claude in conversational AI tools, and a controversial study revealing AI coding tools may actually slow developers down despite perceived productivity gains.