YC invented the SAFE a decade ago, simplifying the process of raising seed money for founders across the world. Today, we’re releasing a SAFE tool to make it even easier to raise your seed round on SAFEs. And it’s free!

Today we're launching @akonlabs backed by @ycombinator We build GitNexus, an open-source knowledge graph for your repository. Coding agents query it to get exactly the context a task needs symbols, call paths, dependencies instead of grepping around and stuffing half your codebase into a context window. Context is the bottleneck for coding agents. Bigger windows haven't fixed it. Buyers of agents are paying more per token to get worse code, and the agents themselves have no map of the repo they're editing. We're fixing this. On DeepSWE, the same agent with GitNexus vs. without: 1. 37% → 68.4% of tasks passed 2. Half the cost per fix 3. On the hardest tasks, the gap more than triples One stack: knowledge graph, coding agent, automated PR review. Managed, self-hosted, or air-gapped. 45k GitHub stars. 65k weekly npm downloads. 10+ enterprise deployments. Sign up at:

Hosted the first YC Intern Expo today with 1,400 college students & 52 YC companies hiring interns. Many of YC's best founders started as interns. From the people I met, pretty sure we'll get more from this crowd.

Yesterday we had dinner with @paulg at @ycombinator's office. And it was by far the most information-dense conversation I have had in my life. I do not think there is another place where you can get this much timeless startup advice in one room. Three things I wrote down: 1️⃣10% week-over-week growth is all you need to care about. It's impossible to mess it up as a founder if you maintain this growth for a long enough period. 2️⃣Great startups ride second waves that propel them to be truly huge. Zuckerberg could not have predicted the coming of the iPhone 3 years after he started. Replit grew from 1M to 100s of millions in revenue in one year from vibe coding. Boom Supersonic couldn't have predicted that there will be such demand for their engines to power AI data centres. You cannot time the wave but it's about being alive long enough to catch it. 3️⃣Formidable founders are deeply committed to what they are building. They survive the hard years because giving up is not an option. They eventually get what they want because they keep going when most people would have stopped. I could have listened for another five hours. What an absolute pleasure to have met him in person.

Everyone in robotics is waiting for the ChatGPT moment but … missing the most important part of the ChatGPT playbook. So we left Boston Dynamics and Amazon Robotics to fix that, and start @Shiraz_AI to build robots that learn on the job. When GPT-3 was released, it couldn’t do every task out of the box. But you could show it a few examples in the prompt, and it adapted on the spot. We’re bringing that to robotics. Show our robot a task once, and it adapts on the spot. That’s exactly what our manufacturing customers need, especially when production lines change every week and traditional automation can’t keep up. And we’re already deploying it with our first customers. Have a factory that’s difficult to automate? We’d love to talk.

Introducing the Halluminate Westworld Finance Diligence Bench: 88 problems that put AI agents through a full company acquisition due diligence process. We tested a range of models across several harnesses. Results below. 🧵1/

Base models for robots are improving but the hard part is always the last mile. @enact_labs is the post-training infra that's been missing to get robots to do reliable work in the real world. Very impressed with what @jamesw_stevens and @govindchada have built. Congrats on the launch 🚀

a patient verifies their identity once. in a couple clicks, @hubblehealthai assembles their medical records from across their providers for your AI agent to use. we connect into EHRs and HIEs, and leverage browser and voice agents so you don't have to do any work.

At Startup School this year, we interviewed 8 awesome ML student researchers at their posters during "YCML", our inaugural research symposium. @ycombinator is funding more research-focused companies than ever before, and we were thrilled to see a huge overlap between students interested in research and startups. Check out their awesome work below in thread.

Models aren't the bottleneck anymore, context is. gives agents live web context without developers having to build and maintain infra. Congrats @mynameisyahia on the launch 🚀

Today, we’re proud to launch Buildbox with @ycombinator 🌟 Most AI agents don’t fail where teams are looking. They pass evals, their traces look normal, and the agent says it completed the task. Then the customer asks again, redirects it, or gives up. There’s no obvious error or failed trace…but the user still didn’t get the outcome they wanted. Buildbox finds those hidden agent failures, organizes them by task, and gives teams evidence-backed fixes they can review and ship. If you’re building customer-facing AI agents, let’s chat!

Radiology scheduling is deceptively hard. Booking an appointment means understanding the clinical protocol, machine setup and a dozen other nuances. These live in legacy systems and people's heads. Generic voice AI can't handle it. Avoca's built radiology-specific AI that runs the path from referral through results and they're already live across 250+ clinics. Vertical AI wins when the vertical is genuinely hard.