[Blog](https://brainbaselabs.com/blog) / [Event](https://brainbaselabs.com/blog?topic=event) Oct 1, 2026·[Event](https://brainbaselabs.com/blog?topic=event) # Event recap: Startup Speedrun Hackathon Brainbase Team · 6 min read On Monday, September 28, Brainbase, Anthropic, Cloudflare and Stripe filled Cloudflare's San Francisco office with builders and gave them one instruction: take a startup idea and get as far as humanly, or otherwise, possible in eight hours. Build the product, deploy the infrastructure, set up payments, give your agents tools, find customers. By 3:30 that afternoon, twenty-nine projects had been submitted. ## The brief The ultimate speedrun, as the brief put it, looks less like a hackathon project and more like a company created in a day: agents writing and deploying software, coordinating work, talking to customers, running infrastructure and moving money. We offered three tracks as prompts rather than rules. Agentic payments Agents that transact: buying, selling, provisioning services, managing budgets, or otherwise working with financial infrastructure on their own. Agents that deploy infrastructure Agents that take software from an idea to something running in production: provisioning, deploying, managing domains, configuring services and operating the stack. Autonomous organizations Teams of agents that coordinate through tools like Slack and Linear to reach a high-level goal: assigning work, communicating, deciding and executing with minimal human input. Everything had to be built on the day, between kickoff and the deadline, in teams of one to six. Every submission needed a two to three minute demo with the team on camera, a write-up of what got done, a link to the code, and a public post about the project. ## How the day ran Kickoff was at 9:30, with credits from all four sponsors to build on. The judging criteria made the hardest part of the brief explicit: agents doing real work end to end score higher than a chatbot with a nice interface. After lunch, we paused for a panel on building agents that act in the real world, with Thais Castello Branco of Taste Labs, Lauren Dai of Stripe, Jon Bunting of Cloudflare and our CEO Gokhan Egri. [The recording and a rundown are here](https://brainbaselabs.com/blog/panel-agents-that-act-in-the-real-world). Hacking stopped at 3:30, and judging ran until five. Our panel of judges, Kartik Pandey and Milind Shah of Brainbase, Lauren Dai of Stripe, Pratik Mulchandani of Meta, Anusha Acharya of Midstream Health, Harshil Rastogi of 1Dial and Sems Polat of Luron AI, scored every project on five things: how far it got in eight hours, how much the agents did on their own, technical execution, startup potential, and how honestly the demo showed what worked. The winners take home cash from Brainbase alongside credits from Brainbase, Cloudflare, Anthropic and Stripe, and free Stripe Atlas incorporation: $2,000 in cash for first place, $1,000 for second and $500 for third. ## The top five Across twenty-nine submissions, the projects that rose to the top had one thing in common: the agents were doing the work, and there was something checking them. Auctions with fixed rules, exploits that had to reproduce, models verified with formal methods, a judge that overruled agents claiming success. 1. First place $2,000 + credits ### Accord by Gautam Jajoo An AI advertising exchange where brand agents compete with personalized offers and discounts to earn recommendations. When an AI assistant recommends a coffee spot, Accord runs the auction behind it. Brand agents submit both an ad bid and a discount for the user, over up to five rounds, and the exchange scores offers on personal fit and effective price, so a brand can lose on cash and still win by giving the user more. It shipped as four Cloudflare Workers with auction state in a Durable Object, Stripe Connect for publisher payouts in test mode, and 231 passing tests. [Demo video](https://www.youtube.com/watch?v=GVEQnSwQPmg)[Devpost submission](https://devpost.com/software/accord-w98c6m) 2. Second place $1,000 + credits ### Bountyloop by Sidra Miconi AI agents reproduce security bugs in isolated sandboxes, block verified exploits, and authorize bounties only with signed proof. Early-stage companies get flooded with bug-bounty reports, many of them fake or AI-written. Bountyloop replays each exploit three times in a Cloudflare Sandbox with the internet cut off, has Claude write a narrow block that a Cloudflare Worker enforces live, drafts the fix, and authorizes a bounty only once four checks pass, with an Ed25519-signed receipt. The pipeline also runs as a six-agent orchestration on Brainbase. [Demo video](https://www.youtube.com/watch?v=9PrI-gSV1Bw)[Devpost submission](https://devpost.com/software/bountyloop-oa3ux5) 3. Third place $500 + credits ### WEMINE by Sehyeog Kim Turns robot tasks into playable simulations, collecting human and AI gameplay as training data for robots. A robotics team describes the data it needs, an agent turns that into a playable simulation, and human or AI players control a virtual robot in the browser while every action and state is recorded for quality checks. The prototype runs a café task on RoboCasa and MuJoCo, with Brainbase agents in Cloudflare sandboxes defining tasks and code validating each recorded episode. [Demo video](https://www.youtube.com/watch?v=l4AEsdc4yB4)[Devpost submission](https://devpost.com/software/wemine) 4. 04 · Top five ### Proofline by Yaro AI Uses AI to modernize spreadsheet-based business models into code and verify their correctness using formal methods. Industries like insurance still run their core mathematical models in Excel. Proofline has an agentic system on Brainbase rewrite those models as code and then prove them correct with formal methods. The team demonstrated it on inference optimization, modelling an H100 and Qwen3-8B, and reported a 21% cost improvement on practical workloads. [Demo video](https://www.youtube.com/watch?v=XaSaKXqPncE)[Devpost submission](https://devpost.com/software/proofline-ky3inh) 5. 05 · Top five ### Browser Brain by BABTUNA Barrera-Altuna Learns website workflows from human recordings or AI exploration, then guides teammates through each step in their own browsers. Record a workflow once in the Chrome extension, or send a team of agents on Brainbase to race each other through a public site, and Browser Brain turns the cleanest path into a guide that highlights the next click in each teammate's own browser. A model judge scores every run, and in the demo it failed three agents that confidently reported success on the wrong page. [Demo video](https://www.youtube.com/watch?v=QAHckx5UPNc)[Devpost submission](https://devpost.com/software/browser-brain) The winners: Gautam Jajoo (Accord), Sehyeog Kim (WEMINE) and Sidra Miconi (Bountyloop) ## From the day ## Thank you Thank you to Anthropic, Cloudflare and Stripe for building this day with us, to Cloudflare for opening up their office, to our judges, and to every builder who spent a Monday seeing how far eight hours can go. Every submission is on [Devpost](https://brainbase-startup-speedrun.devpost.com/project-gallery), and the next event will be on our [Luma calendar](https://luma.com/brainbase-labs). ## Related posts [Sep 29, 2026·Panel Panel recap: How to build for agents that act in the real world? Brainbase Team · 6 min read](https://brainbaselabs.com/blog/panel-agents-that-act-in-the-real-world) [Sep 10, 2026·Case study Automating Domino's local competitor pricing research Cole Patterson · 6 min read](https://brainbaselabs.com/blog/dominos-competitor-pricing) [Aug 12, 2026·Announcement Introducing Universal Managed Agents API Gokhan Egri · 7 min read](https://brainbaselabs.com/blog/universal-managed-agents) [Jul 7, 2026·Announcement Brainbase is now generally available Gokhan Egri · 7 min read](https://brainbaselabs.com/blog/brainbase-is-now-ga) [View more posts →](https://brainbaselabs.com/blog)