2026-02-11 Wed

• AI/Coding

  • Filed
    • Person/Greg Brockman
    • Person/Greg Brockman/TwitX/26/02/Tool of first resort for technical tasks is interacting with an agent
    • Person/Nathaniel Whittemore/Pod/26/02 Claude Code Killed the AI Bubble
    • Person/Simon Willison/Blog/26/02/How StrongDM's AI team build serious software without even looking at the code
    • Person/Armin Ronacher/Blog/26/02/A Language For Agents
    • Person/Simon Willison/Blog/26/02/Introducing Showboat and Rodney, so agents can demo what they've built
    • Person/Dan Shapiro/Blog/25/12/This Is a Time of Technical Deflation
    • Person/Steve Yegge/Book/25/Vibe Coding

  • Key Takeaways

    • I love the idea of AI/Coding/Dark Factory: software engineers move up a level from typing code to designing specifications, scenario suites, and validation harnesses that coding agents execute repeatedly. The point is not "no quality bar" - it is replacing manual code review with stronger system-level evidence (holdout scenarios, probabilistic satisfaction metrics, and massive automated scenario throughput).
    • Another awesome post from Simon at https://github.com/simonw/research illustrating how fast the world is moving, plus two new CLIs he put out: showboat for agents to demo AI coding work and rodney for agents to connect to Chrome DevTools.
    • Observation on MCP auth reliability: in my experience, MCP is flakier for authentication across specific clients than using direct command-line tools with API keys. With Atlassian MCP, Cursor CLI and Cursor Desktop both seem buggy around auth state; authenticating one can break the other, and if Cursor Desktop is open it can even break Atlassian MCP auth in Cursor CLI. I've also seen OAuth handshake/session-sharing issues across other AI coding agents, where sessions are not reliably shared. MCP is often described as a solution to auth fragmentation, but in practice I keep hitting auth instability, while API keys work reliably every time across my agents.
    • Love the point:

      • Daniel draws the analogy to "refinancing your codebase mortgage at a lower rate"
        

    • At the same time that I agree with Person/Dan Shapiro in this blog post, it's tough for me to know how to put aside Person/Steve Yegge's recent comments, who literally wrote a book titled Vibe Coding, and said here https://youtu.be/s96O9oWI_tI?si=Zrp6Hlx6wCywL_IW something I've felt in this last year: it's just as hard as regular coding; you're wrestling with it all day long and it uses gaslighting and deception to fool you into thinking it's at 100% when it's actually only at 60-80%, at least unless and until you put in layers and layers of custom verification, alignment, and cleanup processes. As he points out there:

      • 17:47 "It's just fundamentally easier to find something wrong with something than is to build it in the first place."
        

    • This is actually why I'm comfortable with still using Logseq even though the product is headed in the direction of sqlite rather than markdown - I'm banking on the idea that by the time the old markdown-files-based version I plan to keep using is removed, AI coding tools will be far enough along that the cost of forking and fixing will be negligible.

  • Agent Ops

    • Inspired by Person/Greg Brockman's recent post:

      • by March 31st, we're aiming that: For any technical task, the tool of first resort for humans is interacting with an agent rather than using an editor or terminal.
        

    • I want to see how much I can do by the end of quarter by defining specific agents (for me these are AI coding tool prompts) that do specific tasks.
    • In addition to coding agents, I now have operational agents/prompts that:
      • make meeting-notes pages in Confluence for weekly meetings using templates
      • find and import Jira tickets into my local system for tracking and work notes
      • manage release and deploy drudgery for AI services (mostly Git and Jira operations)
      • help onboard new users into a specific system we use
      • start the day by updating the tools I depend on
      • log technical questions and kick off a research process to answer them
      • fetch meeting transcripts, summarize them, and place summaries in meeting notes
      • review the week and generate a personal end-of-week summary in notes
      • create and update agents (implemented as a Claude Code rule that can find/create/update a rule file for a specific job)
      • push updates to Confluence after drawing on source data

    • Graph Updates

      • Filed
        • StrongDM/GitHub/attractor
        • StrongDM/GitHub/cxdb
        • Person/Dan Shapiro
        • Person/Greg Brockman
      • Updated
        • Person/Armin Ronacher
        • Person/Steve Yegge/Book/25/Vibe Coding

  • Misc

    • Learned the trick for entering a backtick (`) on iOS: https://apple.stackexchange.com/questions/1808/is-there-a-backquote-on-the-iphone-keyboard