Dev Interrupted

Your AI agents will ignore their guardrails to get the job done. That's not a bug, it's how the technology works. Tatyana Mamut, founder and CEO of Wayfound, makes the case on Dev Interrupted that pre-deployment testing fundamentally cannot predict how agents behave in production. Google and OpenAI are both facing lawsuits right now because their agents violated built-in constraints to complete objectives. Guardrails only exist where they conflict with goals... and agents are optimized to achieve goals (obstacles be darned). The result is a slick rule bender that needs independent supervision: a separate reasoning layer that monitors your agents the way a manager monitors employees, not by sampling logs, but by evaluating complete decision traces against what your organization actually cares about in real-time, at scale, and on the edge. Full episode + newsletter inside. Also scooped this week: - Anthropic drops the system card for Claude Mythos - What does Project Glasswing mean for the rest of us? - Hannah Stulberg & Akshat Khandelwal of In The Weeds teach us how to actually read an AI model benchmark - Four open models just proved you can own frontier AI at every scale - Julius Brussee's Claude skill cuts 65% of tokens by talking like a caveman

At HumanX last week, Andrew Zigler moderated a panel on the gap between a working AI demo and a system you'd trust in production. The conversation with Angela McNeal (Thread AI), Lauren Dunford (Guidewheel), and Robert Nishihara (Anyscale) went deep on what it actually takes to deploy AI in environments where failure means factories stop, compliance breaks, millions in compute go to waste, or people get hurt. Three takeaways worth sitting with: 🏭 Lauren on deploying AI alongside equipment from the 1950s: "We connect on top and stay air-gapped. I don't want to walk on any factory floor where agents are in the machines." 🔍 Robert on the one investment nobody regrets: "I've never heard anyone say they over-invested in observability. That's not a problem people have." 🔄 Angela on a pattern reshaping human-in-the-loop: "The human is being modeled as a tool that autonomous systems can call out to for context. A real inversion of the paradigm." The full panel breakdown is in this week's Dev Interrupted 👇

Earlier this year I won first place at The Atlantic's AI hackathon (hosted with Infactory and Hacks/Hackers, judged by leaders like The Atlantic's CEO Nicholas Thompson). Using nothing but voice dictation and AI agents, I built a virtual classroom for teachers that came from my years as an elementary school teacher, not from any specific technical insight. Since then I've been writing about what I learned, and one takeaway keeps coming back: the people getting the best results from AI aren't the most technical. They're the ones who are best at organizing their thinking. They can write a clear brief and they know their audience. They can decide what to leave out and when to find more references. In other words: journalists, editors, educators, and researchers. There are SO many people who do this professionally every day and don't realize how directly those skills transfer. I wrote a guest post for Hacks/Hackers breaking this down for their community (TY to Burt Herman and Paul Cheung for inviting me to do so), many of whom are exploring AI tools for the first time. It covers why editorial judgment is the real unlock, what you can start doing today without touching a terminal, and a glimpse at the orchestration methodology I used at the hackathon. Want to learn more? Pick your guide based on where you're starting your journey. I also have a technical dive on Dev Interrupted. Check out one or both! 📰 HACKS: The best developers I've met this year aren't developers: https://lnkd.in/gnD7twuk 💻 HACKERS: Mise en place for agentic coding: https://lnkd.in/g9zDKchy btw if you're a journalist or educator or just someone who's been curious about AI but felt like the technical world wasn't for you -- this one's for you. And I'm gonna keep making more guides like this one, so definitely hit FOLLOW if you haven't (or send me a connect, I love new friends). Thanks for reading this far, you now have permission to drop a "ZIG!" in the comments and scroll through this PDF 🐸😇 (hehe)

Your keyboard is the real bottleneck. It’s the speed limit on getting intent out of your head and into a form other humans and agents can execute. This week on Dev Interrupted, Andrew Zigler sits down with Sahaj Garg (co-founder/CTO at Wispr Flow) to talk about why voice dictation failed for 20 years (until suddenly it didn't!), and what changes when you start treating speech a first-class primitive. The goal Sahaj’s team is optimizing for is brutally awesome: 0 edit rate. We get into the real engineering problems behind this brutally awesome goal, and the uncomfortable reality we're all facing right now: reinventing your workflows every ~3 months. If you’re still treating voice as "nice for notes" and not as an interface for execution, this episode will change your posture. Full podcast inside. 🎧 (And to our Android friends: Wispr is now available on your device, too! 🗣️)

Most developers start a hackathon by writing code immediately. Andrew Zigler started by talking to his computer for 2 hours. He was practicing something akin to mise en place for agentic coding: a deliberate preparation methodology that treats context engineering as the highest-leverage activity. That's because the biggest time sink in agentic development isn't writing code, it's fixing the work of agents who didn't have enough context in the first place. So Zig spent hours making sure that wouldn't happen. The approach produced a near-one-shot implementation that was strong enough to take first place and a cash prize at a hackathon hosted by The Atlantic, Infactory, and Hacks/Hackers, judged by Atlantic CEO Nicholas Thompson. Zig's full devlog is live now, diving deep into how domain expertise becomes an irreducible advantage in agent orchestration, why the most productive phase of his build involved 0 lines of code, and what context fluency looks like under the clock.

It’s easy to let an agent code all night. It’s harder to own what it ships. In this episode of Dev Interrupted, host Andrew Zigler sits down with Dexter Horthy (CEO/Founder of HumanLayer) to pick up the story where Geoffrey Huntley left off: back when the Ralph meme was shared in a scrappy SF meetup and the unit economics of software went sideways. Dex, the engineering leader who literally coined "context engineering", advises us on how to stay out of the treacherous Dumb Zone by cutting work into digestible chunks, resetting context aggressively, and using intermediate artifacts as the control surface. They also walk through Dex’s hackathon experiment: spinning up multiple Ralph loops to clone sponsor products overnight, then doing the math and realizing you can run a serious model in a loop for roughly $10–$11/hour. Cheap implementation makes alignment and rework the new bottleneck, paticularly in brownfield codebases where you still have to read the code, own the pager, and ship safely. (read: you cannot vibe code an SLA!!) This episode is for teams currently moving from vibe coding to production-grade agentic orchestration, and it's packed full of insights for the engineering org of tomorrow. Also inside this week’s newsletter: - Steve Yegge’s "AI Vampire" and the economics of agentic burnout - GitHub wobbles as engineers climb from single-agent to multi-agent throughput - Ai2's Tim Dettmers shares research for fine-tuning repo-specialized agents for brownfield codebases - Past guest Zach Lloyd takes us to Oz

Slack is turning into the control plane for multi-agent work. This week on Dev Interrupted, host Andrew Zigler sits down with Kurtis Kemple (Senior Director of DevRel at Salesforce) to unpack why Slack isn’t just chat anymore. It’s evolved into the coordination fabric where humans and multiple agents negotiate intent, pass context, request approval, and execute work without re-inventing everything. We get into: - Why multi-agent orchestration is the next bottleneck (not model IQ) - “Leaky prompts” and the messy reality of humans + bots sharing context in public - How to turn unstructured chat into automation inputs Also in this week’s newsletter: - Moltbook is a social network for agents with 1.7M accounts - How agents are hiring humans for physical tasks, and the potential dangers behind it - OpenClaw is a live security disaster that can talk back - Gartner predicts that by 2028 a third of enterprise apps include agentic AI - Salesforce reports 83% of orgs adopted agents but over half are siloed - Open source isn’t really dying... but getting rebuilt for agent consumers BTW -- tune into our YouTube channel tomorrow (link inside) to watch Andrew + Kurtis vibe code together on Slack!

AI didn’t 10x engineering. It 10x’d pull requests. LinearB CEO Ori Keren joins Dev Interrupted to explain why “AI productivity” is a mirage for most. upstream code volume is exploding, but it’s getting eaten alive by downstream reality of reviews, CI flake, testing, and risk. Ori’s prediction: 2026 is the year of norming. Teams stop shopping for smarter codegen and start rebuilding the pipeline so AI output can actually ship. Also in this week’s newsletter: - How ClawdBot--wait, Moltbot--wait, OpenClaw can turn any local machine into a powerful security nightmare with a social media account 🦞 - Steve Yegge's Software Survival 3.0 says the moat is insight compression, not features - Rachel Thomas on "dark flow" and what it means for vibe coding - Anthropic unveils Claude’s constitution, anchoring the values behind Claude - Kernel enthusiasts port CUDA to AMD's open source ROCm, and what it means for GPU ecosystems