Weekly Links (Jun 29 - Jul 05)

This week was rich in reflections on the impact of LLMs on software development, from practical assessments of the current state to strategies for making them more reliable. We also highlight important discussions about resilient architecture and the value of expert generalists.

Coding agents have crossed a chasm

Coding agents have reached a crucial inflection point. The author describes how these tools have evolved from experimental curiosities to indispensable daily assistants. Claude Code and OpenAI Codex now complete entire tasks autonomously, fundamentally changing workflows - from creating personal tools to fixing complex production bugs.

Software engineering with LLMs in 2025: temperature check

Gergely Orosz presents the results of his annual research on AI use in development. He spoke with engineers at AI startups (Anthropic, Cursor), Big Tech (Google, Amazon), and independent developers like Kent Beck and Simon Willison. The consensus: tools are spreading rapidly, but we're still figuring out exactly how they will transform software engineering.

From idea to PR: A guide to GitHub Copilot's agentic workflows

GitHub details how Copilot has evolved into complete autonomous workflows - from initial idea to ready pull request. The article demonstrates in practice how to use Copilot Workspace to transform an issue into functional code, including planning, implementation, and testing, representing a fundamental shift from assistant to autonomous agent.

Reliability for unreliable LLMs

Stack Overflow explores how to add determinism to systems based on non-deterministic LLMs. The article covers essential techniques: input/output sanitization (filtering jailbreaks, PII, and toxic content), enhanced observability (detailed logging of prompts and responses), using "LLM-as-judge" for validation, and deterministic workflow orchestration. Includes insights from leaders at IBM, Salesforce, and Vectara.

Tool Calling with Local LLMs: A Practical Evaluation

Docker systematically evaluates the performance of local LLMs for tool calling, testing models like Llama, Qwen, and Mistral. Results show that smaller, optimized models (like Qwen 2.5 with 7B parameters) can match or exceed larger models in specific tasks. Crucial for companies needing on-premise solutions for privacy, cost, or latency concerns. The article includes detailed benchmarks and reproducible code.

The principles of extreme fault tolerance

PlanetScale shares the principles and processes that guarantee their extreme reliability: isolation (physically and logically independent parts), redundancy (multiple isolated copies), and static stability (continue operating with last known good state). The article details how these principles translate into architecture and operational processes, including weekly failovers on all client databases.

Tools: Code Is All You Need

Armin Ronacher (Flask creator) questions the hype around MCP (Model Context Protocol) and argues that simple, well-structured code is more effective. He demonstrates how he transformed his entire blog from reStructuredText to Markdown using LLMs to generate reusable Python scripts, rather than relying on complex integrations. The message: automate with code you can understand, debug, and run 100 times.

Why Organizations Need Expert Generalists

Martin Fowler advocates for the value of "expert generalists" - T-shaped professionals with depth in multiple areas and breadth across many others. The article emphasizes that these professionals focus on enduring domains (distributed systems, cloud-native architecture) rather than specific tools. They're essential for connecting organizational silos, translating between technical and business teams, and leading complex transformations.

🎥 Talks and Presentations

AI-Driven Code Refactoring: Improving Legacy Codebases Automatically - Jorrik Klijnsma

Jul 02, 2025 ⸱ 59m

Kotlin Clean Architecture for Serverless: Business Logic You Can Take Anywhere | Elena van Engelen

Jul 02, 2025 ⸱ 33m 44s

Resilient by Design - Chris Ayers - NDC Oslo 2025

Jul 04, 2025 ⸱ 54m 21s

Claude Code & the evolution of agentic coding — Boris Cherny, Anthropic

Jul 04, 2025 ⸱ 18m 12s