GitHub shipped gh skill last week with the open Agent Skills spec: packaged instructions, scripts, and context that work across Copilot, Claude Code, Cursor, Codex, and Gemini CLI.
Portable agent tooling is finally here.
GitHub's new Copilot CLI "Rubber Duck" uses a second model from a different AI family to review the first one's work. We reinvented pair programming. This time neither pair can grab you coffee.
AI-generated lines of code in 2026 is lines of code in 2006. We already know how this story ends. Measure outcomes, not effort.
Lightrun's 2026 State of AI-Powered Engineering Report (April 14, 2026) -- 43% of AI-generated code fails in production after passing QA and staging. 88% of companies require 2-3 manual redeploy cycles to confirm AI-generated fixes work. Faster is not always better.
75% of startups use Claude Code. 56% of enterprises still default to GitHub Copilot. This is not a tools story. It is a risk tolerance story. And it will determine who ships faster over the next three years.
You used AI to ship twice as many features this quarter. You didn't double your test coverage. You didn't double your QA capacity. You didn't double your monitoring. Congratulations, you've doubled your risk surface and called it productivity.
Software engineering jobs are up 30% in 2026. The AI apocalypse was rescheduled again. Though "up 30%" and "same jobs" aren't the same sentence. The roles opening aren't the roles closing.
GitHub will now train AI models on Copilot data from Free, Pro, and Pro+ users. Opted in by default. The tool you use to write code is learning from how you write code, whether you agreed to it or not. How many devs will actually go change that setting?
Claude Code is now the most-loved dev tool at 46%. GitHub Copilot, which once owned this category, sits at 9%. The tool that wins isn't the one with the best marketing. It's the one that removes the friction engineers hate most. Turns out devs prefer tools that actually think.
AI made your team write 59% more code last year. Your release frequency didn't move.
You don't have a productivity problem. You have a delivery pipeline that wasn't built for this volume.
More code is not more software.
Everyone's fired up about GPT-5.4 crossing the human OSWorld baseline.
I want to be excited. I'm trying. But OSWorld tests defined tasks with known endpoints.
Most of what I actually spend my day on does not have those properties.
DORA replaced their performance tiers with 7 team archetypes this year.
One of them is called "Constrained by Process."
I'm staring at my calendar right now trying to figure out which recurring meeting I run that made the list.
I'm very excited about this announcement and proud to be a part of the team that made this happen! Aiwyn Tax is officially available inside Claude!
claude.com/connectors/a...
We spent 2024 asking "will AI replace developers?" We spent 2025 proving it wouldn't. Now in 2026 we're learning the real question: can your organization absorb the speed AI enables? This might be the hardest problem to solve yet.
The engineers thriving right now aren't the fastest coders. They're the ones who know when the model is confidently wrong. That's not something you learn from a tool. It's a judgment skill built through years of watching things break in production.
GitHub just launched Agent HQ. You can now run Claude, Codex, and Copilot side by side inside your repo. Microsoft is literally selling access to its competitors' AI inside its own platform. When the platform play matters more than the model play, you know the model war is over.
The 2025 DORA report confirms what good engineering leaders already knew: AI doesn't make bad teams good. It makes good teams better and exposes cracks in everyone else. AI is a multiplier, not a miracle.
www.infoq.com/news/2026/03...
AI-native startups are reportedly operating at 5.5x the efficiency of traditional teams. But efficiency doing what? Shipping faster doesn't matter if you're shipping the wrong thing faster. Strategy still can't be automated.
Anthropic's internal memo says junior engineering roles are "dubious." They're right about the tasks. They're wrong about the conclusion. Those junior tasks are how senior engineers learned to think. We haven't built the replacement yet.
Apple is launching 5 products this week.
The White House is meeting with tech companies to ask who's paying the electric bill for all these AI data centers.
One of these stories will get 100x the coverage. The other one matters more.
Glenn Thimmes
@glennthimmes.bsky.social
· now
Do you see this more as a hardware limitation or space where architectural innovation will impact. I get it's both but I wonder which will be more impactful.
Spotify's best engineers haven't written a line of code since December. They supervise AI output now. The job title is the same. The job isn't. Are you preparing your team for that shift, or pretending it isn't coming?
ggml.ai joining Hugging Face is the most underrated AI infrastructure news of the year. Local AI just got institutional backing. This matters more than another GPT-5 rumor.
Sprint planning: 2 hours. Sprint review: 30 minutes. Are we spending more energy predicting work than learning from it?
Story points measure effort. Cycle time and velocity measure speed. Neither measures whether the customer cared. Choose your scoreboard wisely.
