Speed is the most important moat in tech. Yesterday I said every vertical AI company will fail. It's because models can do things today that weren't possible 3 months ago. Companies can make a lot of money building on that edge. Adding harnessing, evals, application logic, and other guardrails will allow you to live on the cutting edge. But in a few years, models will eat all of that harnessing, and any differentiation you've created is gone. So as everyone discusses moats, I think the most important one is the build a business that gets better with the models. Meaning that you ride that cutting edge -- which requires speed. The comfort of "focus" is gone in a world where everyone has to do everything. You used to be able to build a moat around complexity, and complexity compounded with focus. That will kill you in 2026+

Vertical AI startups are all going to fail. It shows a fundamental misunderstanding of the technology and the pace of progress. Remember Jasper AI? When they were founded, they were on the cutting edge, so they were valuable. But 3 years later and what WAS the cutting edge is no longer even a full feature of an otherwise sellable product. The way to build a company in 2026 is to orient around the cutting edge. Build a machine to stay on the cutting edge -- or better yet, to push it further. Every decision we make at @tenex_labs is to optimize toward that.

AI made CJ Hess faster—and hourly billing meant he earned less for the same work. Tenex built an engineering organization around paying for output instead. @businessbarista and @seejayhess of @tenex_labs join @swyx at @aiDotEngineer New York to discuss story points, agent workflows, enterprise AI, and what happens when you build your own software. 0:00 Cold open 3:17 Paying engineers for outcomes 7:20 Inside MetaHarness 19:29 The enterprise AI adoption gap 23:22 You can’t automate a process that doesn’t exist 24:42 The debate over vertical AI 30:32 Kill My SaaS: build vs. buy 35:23 AI and jobs 38:18 New York’s AI ecosystem Full writeup, show notes and audio: Watch on YouTube:

AI Agent Sprints have blown up since we launched them a few weeks ago. Have received interest from hundreds of enterprises including: - Tesla - Airbus - SharkNinja - Amazon - Palo Alto Networks - Samsung We have a few spots left, before we close the October cohort. These 5-day, free intensive bootcamps for enterprise leaders are the perfect way to FINALLY get agents to production. At @tenex_labs, we've helped dozens of midmarket cos, F2000 enterprises, and PE portcos deploy agents into production. With bootcamps, we're lowering the friction of experiencing the power of an agent within your company's workflows to ~zero. 3 phases of the sprint: 1) Find Opportunities - We interview the people closest to the work to identify where agents could have the biggest impact. 2) Map Workflows - We map workflows step by step, including the systems involved and expected business impact. 3) Prototype Agents - We build prototypes on your data, demonstrate the results, and recommend what to take live first. Sprints are led by our world-class FDEs & strategists, on the ground, working with your teams and systems. What do you get? 1) A clear AI roadmap with ROI per opportunity 2) Up to 8 mapped opportunities 3) Up to 4 working prototypes Who are agent bootcamps for & how much do they cost? VP-level & up decision makers at midmarket & enterprise companies doing >$100m in revenue or funding. Bootcamps are free. They're an opportunity to see what AI agents can do for you before committing. Why are we launching this now? Many companies are in the messy middle of AI transformation. They've rolled out LLMs companywide, perhaps done some AI enablement work, coding agents are being used widely in engineering, and that's it. The next obvious step in the AI journey is bringing agents to the workflows of an organization, but most companies are getting stuck. Taking agents to production is complex and nuanced and we've built a repeatable process for handling the final mile of applied AI, which allows us to take agents into production in ~2 weeks. How do I sign up? Join the Agent Sprint waitlist here:

If you’re a non-linear thinker, builder or engineer and while reading this you felt an urge to contribute in diffusing this technology positively and safely to people. Please reach out to me, @BrettFranklin_ , @ArmanHezarkhani or @businessbarista with the hardest problem you would like to solve if you’re unleashed with full autonomy. @tenex_labs is NOT a consultancy, we : - bridge the gap between people and ethical AI - help organizations transform so they don’t get left behind - have the smartest, kindest humans collaborating to solve critical problems at scale We’re actively looking for bright minds like yourself — so feel free to apply directly: This window of opportunity won’t last long!!

Being adaptable is more important than ever / might become the only thing that matters. You must let go of the past and embrace the future cause we're already moving too fast to stay on the bleeding edge of everything. And we're just getting started

Funniest thing I heard this week was a member of the team being surprised by how cool the people we met at last night’s NYC meetup were compared to typical tech meetups. “There were people drinking martinis!”

WTF is a smoke test? This is a question software developers working with coding agents have probably asked themselves at least once. The whole concept of "smoke testing" has become an inside joke, as it is usually a clear indicator of AI slop code. Things like “synthetic smoke tests” or "Canonical UI smokes", are dead giveaways. My coworker, @romanvellante, and I kept coming back to the same question though: Why do smoke tests suddenly seem to be everywhere? So we dug into it, and wrote this article covering: 1. Where smoke tests came from 2. Why coding agents keep creating them 3. When they are useful, and when they become clutter 4. How to keep repositories that coding agents work in maintainable Hope everyone finds this piece helpful!

Opus 5.5 and Astra have me rethinking the baseline again. A few thoughts that I feel like went from an idea to concrete in my mind over the past few months: 1/ Delete tests - the models love to slop up tests and check that “abc” == “abc” all over the place. A lot of tests restate the code instead of test the code. This coupled with the fact that the models just write correct code these days has me feeling like unit tests are now just a waste of tokens. 2/ PRs and code review are a ceremony at this point - maybe on super critical pieces like payments, auth, etc. it’s still worth going line by line. Otherwise, have the model write the code, have the model test the code, have models review the code, and have the model drive computer use to prove that it works end to end. Stolen from @thorstenball , “give me irrefutable proof this works end to end” often I ask for this in the form of a video and then I review demos and poke around myself. 3/ It’s orbing time - Of course everything is moving to cloud. I have discussed @AmpCode a lot. You can see this pattern with Codex cloud from dev day. Moving your work to the cloud is a second unlock that is not about using the models well. 4/ The bar is higher - Motion design videos just got commoditized. You need to innovate where the models struggle (not much room anymore) or else you’re ngmi. 5/ Teammate, not tool - The models these days can triage and monitor. I don’t push releases without a thread that’s watching it, reading logs, and writing some doc. At worst case, I build up context about what actually happens in production. Best case, I get a hot fix automatically. 6/ Hillclimb - Overused word rn, but performance, optimization, and sweeping refactors are no longer scary. If you close the loop on task → validation and it can be proven on a computer, it’s solved. I think some of us were early to jump to these conclusions with Sonnet 3.5, but they’re looking correct. At the same time, I have no idea where this goes from here.