John Tan

We backed MuleSoft back in 2006 to build the leading platform for enterprise adoption of APIs. WunderGraph is the next MuleSoft.

20

87% of YC companies are enterprise. Now it makes complete sense that AI is starting in enterprise. The models are expensive to run, and enterprise lets you automate workflows and vertically integrate. But to reach billions of people in their daily lives. The real prize is still consumer. A lot of founders are scared of consumer right now. At any point, OpenAI or Google can vertically integrate and kill your company. But you can make a similar argument for enterprise applications. The other part is that consumer is genuinely hard. Distribution is more mature, the App Store feels saturated, and there aren’t obvious new channels. At the same time, we are in the midst of a revolution. Interfacing with the world will be more than just a handful of chatbots. If you pay attention to daily life, to your parents, your kids, anyone close to you, you’ll notice how many small, annoying things still exist. Yes, the App Store is mature. But the top apps right now are AI apps. If something is truly great, innovative, or fun, people will go out of their way to find it. Menlo Ventures analyzed routine tasks of US adults. The Takeaway: Fewer than one in five people actually use AI for their day-to-day tasks. Most people stick with whatever system they already have, even if it’s annoying. At Hands, we're building with these 3 rules in mind: 1. Meet people where they already are 2. Make it easy to try. 3. Prove the product's value almost immediately

4

Developers shouldn’t be the bottleneck for issuing refunds, extending trials, or adding customer credits. Happy to share that Custom actions are now live in bricks.sh. Extend your admin panel beyond CRUD operations—and empower CS, ops, and business teams to trigger Supabase Edge Functions directly from the interface. No code. No Slack pings. No context-switching. Here’s how different teams will use it: 1/ Customer support: trigger refunds, add credits, and extend free trials via safe edge functions. 2/ Operations: bulk-update orders, shipments, and subscriptions from one panel. 3/ Finance: run payouts, invoices, and reporting workflows on demand. Devs write the function once. Business teams run it safely―whenever they need. You stay in control, they stay unblocked. In short, how internal tools should work in 2026. --- You can connect your Supabase database and generate your admin panel at https://brcks.sh/fl2-d

30

1 Comment

Garry Tan, president and CEO of Y Combinator, recently posted about why Ruby on Rails plus Claude Code is a "crazy unlock." Rails is designed around strong conventions, predictable patterns and expressive DSLs. That combination means an LLM has far less ambiguity to fight through before it can be useful. As he put it: "Rails was designed for people who love syntactic sugar, and LLMs are sugar fiends." And it got me thinking about Django. Django has some convention. The framework will tell you roughly where your models and views and URLs live. What it won't tell you is where your business logic goes. Or how your views and models should be structured. Those questions get answered differently by every team, in every project. The thing about those decisions is that you don't really know if they were good ones until years later. A pattern can feel perfectly sensible in year one and leave you maintaining a small archaeology project by year five. That experience is what pushed me to write Django RAPID Architecture. RAPID adds convention where Django leaves things unsaid. It gives you a clear horizontal structure: data, interfaces, readers, actions. It makes it obvious where different kinds of code belong. And with django-readers, the specs layer gives you a compact, expressive way to describe what data a view needs and how it should be shaped: something close to a DSL. Well, it turns out that Garry was right. When the codebase has already made the structural decisions, a model doesn't have to reverse-engineer your team's habits before it can be useful. With a bit of guidance, the shape of the thing is immediately visible. And LLMs love it. I've seen agents one-shot large, complex features, and the code is exactly as I would've written it. That's the deeper point in Garry's post. It's not really about Rails specifically. The most AI-native codebases are the ones that are already legible, where the important decisions have been baked into the structure itself. So RAPID isn't just for humans, it's for agents too. If you want help moving your Django codebase in the direction of these huge AI productivity gains, get in touch: https://lnkd.in/eiRtAepq And if you just want to learn more about RAPID: https://lnkd.in/ejkF7DMB

100

18 Comments

Terminal Use (YC W26): Vercel for background agents. Terminal Use is hosting infrastructure for background agents that makes harness engineering easy. https://lnkd.in/gv99wN2t

51

2 Comments

The big labs are expansive in their product ambition, especially since foundation models have largely improved in lockstep - in order to compete with them you have to do things they won’t which are: - building a very rich software ecosystem around a primitive - orchestration across multiple models - going insanely deep on product and growth for a narrow vertical domain

58

3 Comments

"Lithic, a card processing platform, now integrates with Visa's ALM via the VCPE API for enhanced card program delivery and economics." See this and many other announcements in this week's issue of the API Changelog newsletter: https://lnkd.in/dAPJm-fs

4

Mike Merrill and I are excited to announce the next chapter of Terminal-Bench with two releases: 1. Harbor, a new package for running sandboxed agent rollouts at scale 2. Terminal-Bench 2.0, a harder version of Terminal-Bench with increased verification Harbor is the package we wish we had had while making Terminal-Bench. It’s for agent, model, and benchmark developers and researchers who want to evaluate and improve agents and models. Just a few of the features I love about Harbor: - Evaluate any agent that can be installed and run autonomously - Scale up to thousands of concurrent containers using providers like Daytona and Modal - Generate rollouts for SFT data gen and RL - Create your own benchmarks or use existing ones like SWE-Bench Verified Harbor is also the official harness for Terminal-Bench 2.0. We used it to run tens of thousands of experiments in containerized environments while developing the latest version of the benchmark So why Terminal-Bench 2.0? We always knew that as model capabilities increased, we’d need to keep Terminal-Bench up to date with frontier capabilities. Terminal-Bench 2.0 consists of 89 hard tasks to test these capabilities. We aim to push the frontier with an increased emphasis on task quality. Each task received several hours of human and LM-assisted verification to ensure that tasks are (1) solvable, (2) realistic, and (3) well-specified. We’ll share more on how we did this in our upcoming preprint. At present, Codex CLI with GPT-5 sits at the top of our new leaderboard. Astute Terminal-Bench fans may notice that SOTA performance is comparable to TB1.0 despite our claim that TB2.0 is harder. We believe that this is because task quality is substantially higher in the new benchmark. We have removed several misspecified or impossible tasks – increasing difficulty while maintaining raw performance. Interested in using Terminal-Bench-2.0 or submitting to our new leaderboard? Check out the Harbor docs for more information. It's been a pleasure working alongside our advisors, Ludwig Schmidt, who seeded the idea and guided our development, and Andy Konwinski, who gave us the Laude Institute mandate to "ship your research." Additionally, Terminal-Bench wouldn’t be possible without its community. We’re so thankful to the over 1k members of our Discord who contributed and audited tasks, helped build and beta test Harbor, and made this such a fun project for everyone involved. https://lnkd.in/guZXbPVn https://lnkd.in/g7xqkjK8 https://lnkd.in/gETM-wkx

110

5 Comments