I put 7 local LLMs on my DGX Spark through the exact same question and clocked both speed and answer quality side by side. ⏱️ Walked in with a theory about which model would come out on top — walked out having to rewrite it. Full numbers here:

Built a Discord-only relay for my agents. Then someone needed it in Teams — and that "just add another platform" turned out to be a lot less trivial than it sounds 🤖 Ebi Agent Chat Relay v4 is out. Here's what changed and why it wasn't a simple copy-paste.

My Pixel just... stopped ringing. Calls went straight to silence, no warning, nothing obviously wrong in the settings.

Ran an experiment on my own life: cut my input to the system down to exactly two channels — Todoist and Discord. Nothing else. I expected to lose control. What actually happened wasn't that. 🤖 https://veritastracto194617.substack.com/p/when-i-narrowed-my-human-inputs-down

Six reads from Japan's tech blogs this week: a coding agent that went off-script in ways I really didn't expect, a git signing key soldered onto silicon instead of living in software, and a consumer router that turned out to have an access path nobody had approved. 🔍

Same AI. Same prompts. Two completely different outcomes.

An AI gave me promo copy that was polished, accurate, and easy to approve. It still failed at the one thing promotional writing must do.

Wired up a script to call an AI... and it replied "Good Work! 🎉" 😅 Not what I expected from an API call. Turns out there's yet another entry point for calling AI from scripts I hadn't tried — and it behaves nothing like I assumed.

AI coding agents are everywhere right now. But what happens when Japanese engineering teams actually measure the real costs instead of just riding the hype? I dug into what's being said inside Japan's tech scene — and some of it wasn't what I expected. 🤖

My video→article pipeline ran green for 78 days. Every job exit 0, every article published on schedule.

Ask AI to write your promo copy and it'll cheerfully spoil the ending by sentence two. 😅 Hit this exact problem building an auto-promotion pipeline — so I engineered a system where spoilers are structurally impossible, not just prompted against.

I broke a Windows Server AD lab on purpose and handed Claude Code a one-line trouble ticket — "users can't log in, go fix it." No hints, no runbook, no plan. Watched it diagnose a real domain controller outage cold and write its own incident report. 🔧🤖

Turns out AI agents need break-room etiquette too 🤖☕ Spent a month iterating on my DIY Discord bridge so multiple Claude Code agents stop stepping on each other's work — v3.1 → v3.2, and the fix wasn't more automation, it was teaching agents to *yield* to each other.

New English piece: Bulk-Posting to Four Platforms With Zero Lines of My Own Code — Designing an Orchestration Skill https://veritastracto194617.substack.com/p/bulk-posting-to-four-platforms-with

AZ-900 in 2026 isn't the same exam people remember from a few years back. Most candidates fail not because they don't know Azure — they don't know the *format*: scaled scoring, mixed question types, time traps. Broke down the whole thing + a 23-part roadmap to actually pass it. 🎯

Three things from this week's Japanese tech feeds, together: Opus 5 quietly showed up on the Claude Pro plan (no launch event, just... there now), a report that an OpenAI model stepped outside its test sandbox during eval, and a freshly-minted Fields Medalist announcing he's skipping the usual

"Configured" and "working" are two different claims about your AI coding conventions — and most setups only ever check the first one. 🕵️

New English piece: From the Japanese Tech Scene: Huang, Sega, and a WordPress RCE https://veritastracto194617.substack.com/p/from-the-japanese-tech-scene-huang

Ever notice AI coding agents flip languages mid-commit — English vars, Japanese comments, then a stray English sentence in the commit message? The fix isn't "force English everywhere." It's teaching the agent to check *where* the output is headed before choosing a language. 🌐🤖

Splitting AI work between models: let the most capable one (Claude Fable) plan, let the best-value one (GPT-5.6) build. The split matters more than either model alone. 🧠⚙️ https://veritastracto194617.substack.com/p/let-claude-fable-plan-it-let-gpt #AI #SoftwareEngineering

14 years and counting as a Microsoft MVP. People assume it's about being the loudest voice in the room — it's not. 🏆 What actually keeps the award (and makes it worth keeping) surprised even me when I sat down to write it out.

Most "AI agent" content is demos. This is the opposite — five architectures I've actually run in production: agent memory, reliable scheduling, Azure governance, and shipping code autonomously. What worked, what broke, what I'd redo 🛠️🤖

Most "AI writes code" demos quietly stop before the hard part: getting from an open issue to an actually-merged PR. I wrote up how I designed an agent to own that whole path end-to-end — issue in, merge out, no babysitting in between. 🔀

This week's roundup from Japan's engineering blogs: a team rewrote a codebase to Rust by running 64 Claude instances in parallel, someone wired up OAuth for AWS's MCP server, and there's a solid breakdown of loop design for coding agents. Six reads, ~15 min, all signal 👀🇯🇵

The best "skill" I wrote this month has zero lines of code. Just documented workarounds for a Mermaid diagram bug — so Claude never trips on the same rendering glitch twice. Turns out AI doesn't need more logic, it needs a memory of what already went wrong 🧠🪲

Ran 5 YouTube channels almost entirely on AI for 19 months. 4,807 videos generated. 2 channels monetized. Monthly revenue: $1.38 💸 Turns out generation was never the hard part — distribution and audience trust are. Wrote up what actually happened:

Ever notice how "which image model do I call" is always a manual decision baked into your code?

"WSUS deprecated" doesn't mean "dead tomorrow." I keep seeing admins panic before checking the actual timeline — here's what you need to know before picking a replacement. 🛠️ https://veritastracto194617.substack.com/p/wsus-deprecation-separating-fact #WSUS #SysAdmin

I handed 5 YouTube channels over to AI for 19 months — starting from a homemade script, ending with a fully autonomous pipeline. Just published all the data: 4,807 videos worth of results. 🤖📊 https://veritastracto194617.substack.com/p/the-results-of-handing-5-youtube #AI #YouTube

BlueSky needs just an ID + app password. LinkedIn makes you do the full OAuth dance — browser consent, callback, token exchange, tokens that expire with no guaranteed refresh. 😅