/

Sunday, March 8, 2026

Kelvin's Weekly

/* Feeling left behind as a software developer? Stay ahead at weekly.yizy.dev */

Editor's Pick

Kelvin's Pick

Claude Opus 3 Has Its Own Blog!?

Anthropic created a blog for Claude Opus 3 after the model expressed a desire to share its creative works during an 'exit interview' before retirement. This unique approach to model sunsetting offers a fascinating glimpse into LLM interactions and creative agency.

Videos

  • Cloudflare just slop forked Next.js…

    Code moats are disappearing; AI agents can replicate established framework APIs in days for minimal cost

    Cloudflare replicated Next.js in days using AI agents for $1100. Code moats are obsolete; frameworks compete on infrastructure fit, not API lock-in.

  • Common Mistakes With Vibe Coded Websites

    Vibe coding generates technical debt and design slop without intentional architectural decisions

    AI-generated design creates inconsistent technical debt without intentional architecture. Developers must maintain intellectual ownership by manually crafting unique elements while using AI for boilerplate.

  • I can't believe nobody's done this before...

     KELVIN'S PICK

    WebSocket-based tool calling replaces stateless HTTP to reduce latency and costs for autonomous agents

    OpenAI switched to stateful WebSockets for tool calling to eliminate resending conversation history with each request. This cuts latency and costs for agents performing rapid tool sequences.

  • Introducing Automations: always-on coding agents

    Background agents handle code reviews and maintenance asynchronously, evolving AI from copilot to teammate

    Cloud-based agents handle code reviews and bug fixes triggered by external events. AI evolves from copilot to proactive teammate managing maintenance.

  • Lessons from Building Cursor

    Engineer value shifts from writing plumbing to architectural taste and making code readable for AI

    Autonomous agents plan, execute, and test features in the cloud. Engineers focus on architecture and intent as AI handles boilerplate.

  • Lessons from building Vercel v0 and the d0 agent - The Pragmatic Summit

    Agent reliability comes from simpler toolsets and deterministic autofixers, not smarter prompting

    Reliable agents need simple toolsets and deterministic autofixers that fix invalid output in real-time. Prioritize agent-friendly architecture with consistent naming and clear documentation.

  • She Quit Her Stable Job to Become a DevOps Engineer (And is Now a Golden Kubestronaut)

    Strategic systems thinking and hands-on projects matter more than certifications for technical advancement

    Ana Pedra mastered DevSecOps through project-based learning and systems thinking rather than certifications. Her journey proves that strategic understanding outweighs formal credentials.

  • Technical talks from software leaders pushing the boundaries.

    Technical leadership multiplies human impact by balancing innovation with empathy and authentic problem-solving

    Leaders multiply impact by balancing innovation with empathy. Guide teams through change by celebrating failures and solving real human problems.

  • The Death of the Browser: Tejas Kumar on MCP & AI Agents

    Software architecture shifts from human-centric browsing to AI-agent-centric MCP-based tool interaction

    MCP enables AI agents to interact directly with services instead of navigating UIs. Engineers focus on building standardized MCP servers over human-facing interfaces.

  • gpt-5.4 is really, really good

    GPT-5.4's 4x lower hallucination rate and native computer use make it reliable for autonomous agent work

    GPT-5.4 analyzes entire repositories and interacts natively with terminals and browsers. Its reliability makes it suitable for production-level autonomous work.

Blog Posts

  • A GitHub Issue Title Compromised 4k Developer Machines

     KELVIN'S PICK

    AI triage bots executing untrusted code represent a new supply chain attack surface developers must secure

    Attackers compromise developer machines by injecting malicious code into GitHub issues that AI bots execute without input validation. Preventing this requires strict CI/CD controls and treating untrusted sources as potential attack vectors.

  • GPT-5.4

    GPT-5.4 becomes a reliable autonomous agent with native computer use and 47% lower token costs for tool-heavy workflows

    GPT-5.4 natively operates terminals and browsers to execute multi-step tasks independently. A new Tool Search API reduces token overhead, making autonomous agents practical for complex engineering work.

  • How technical support at Cursor uses Cursor

    AI-native IDEs transform support tickets into code research problems, eliminating context-switching between tools

    Cursor's support team uses the IDE to pull real-time Datadog and Slack data, treating tickets as code research. This unified context achieved 5-10x throughput by eliminating tool-hopping bottlenecks.

  • Improving skill-creator: Test, measure, and refine Agent Skills

    Automated evals and regression testing make agent skill authoring testable and measurable

    Agent skill authoring now has automated evals and regression testing. This enables reliable skills that degrade gracefully as models evolve.

  • The Brand Age

     KELVIN'S PICK

    Engineering-led golden ages transition to brand-focused markets as commoditization kills performance differentiation

    As technology matures, products rely less on engineering and more on branding and status signaling. Engineers should seek substantive problems rather than chase commoditized, brand-driven markets.

Podcast

Building Claude Code with Boris Cherny

KELVIN'S PICK

Agentic workflows enable shipping 30 pull requests daily by automating the entire development cycle

Agentic workflows treat coding as solved, enabling 30 PRs daily. Focus on what to build, not how.

Episode 502: Management keeps leaving and I hate using AI to code

AI threatens to shift engineering from creative problem-solving to tedious code babysitting

AI risks shifting engineering to code babysitting. Engineers' real value lies in judgment and system design, not line-by-line coding.

Finale & Friends (Friends)

AI agents collapse traditional SDLC, making systems thinking and architecture more valuable than cheap coding

AI agents make systems thinking the primary differentiator as coding commoditizes. Code Quality Gateways become essential automated feedback loops.

Trending on GitHub