The Agents That Scaleups Built: Stories From AI-First Builders

Once upon a time, AI was a nice-to-have. Now it's the thing scaleups are built on. During Lisbon AI Week 2026, join real practitioners from four fast-growing companies for an evening of candid stories about building AI-first, at speed. They’ll share what shipped, what broke, and what they'd do differently. Lineup: Kianna Love and Jake Duffy (Paddle), Quentin Churet (Alpic), Tiago Zien-Mendes (Speakeasy), and Faateh Dhillon (Dust) Location: IDEA Spaces - São Sebastião From products designed to be promptable from day one, to an agent that supports a daily habit, to agents trusted enough to act with real autonomy, to AI that's shared across a whole team, no two stories are the same. They’re all chapters in the same tale: what it takes to build with AI right now. Grab your spot, get a drink, and let the stories unfold. Meet the speakers Quentin Churet, Alpic The Owl Doesn't Know Me - Replacing Duolingo with an MCP App and the agent that already knows my life I hit B1 in Spanish and stopped. Duolingo had taken me as far as a fixed curriculum can : it doesn't know that I'm currently reading a book about Pablo Escobar's life, or that my wife is the reason I'm learning at all. Meanwhile the agent I talk all day knows all of that - and forgets it by tomorrow. Agents have context, they have no record. So I built Gordito, an MCP App that gives my agent memory of my Spanish. It stores what I got wrong, not just what I looked up. The server owns when I see something again. The agent owns what it looks like - as an interactive quiz, directly inside the conversation as an MCP App. I'll show the build, the demo and what I've learned so far. Faateh Dhillon, Dust The Interruption Tax - Reducing team interruptions with AI agents “Who owns this?” “Have we seen this before?” “Can someone explain how this works?” They’re small questions, but answering them pulls people away from their work. The information usually exists somewhere: in a Slack thread, a document, a past decision, or a colleague’s head. Finding it becomes someone else’s job. This talk explores how we approach that problem at Dust with agents a whole team can use. I’ll walk through a concrete workflow: how an agent gathers company context, checks live systems, and helps resolve a request before someone has to interrupt a colleague. I’ll also cover where that approach falls short, when a person still needs to step in, and what the agent should bring them so they don’t have to start the investigation again. Through a demo and the design choices behind it, we’ll look at what makes an agent useful beyond a one-person chat: shared knowledge, access to tools, and clear boundaries. The goal isn’t to stop people talking. It’s to make fewer conversations start with “sorry to interrupt.” Kianna Love and Jake Duffy, Paddle Is Your Platform Promptable? What does it actually take for an API-first platform to be genuinely promptable — usable by an AI agent that's never seen it before, with no prior knowledge, using only public docs as context? Paddle has spent years building API-first, treating every capability as an API before it ever becomes a UI. That discipline turned out to be exactly what was needed when Lovable came knocking. Through the partnership, developers can now build, go live, and manage transactions — subscriptions, refunds, and all — entirely inside Lovable, without ever returning to Paddle. What used to take days or weeks of integration work, an AI agent can now complete in 10 minutes. This talk breaks down what it actually takes to get a platform there — and what changes, technically and organisationally, once AI agents become one of your platform's primary users. Tiago Zien-Mendes, Speakeasy Permission to Act In a world where AI agents are moving beyond chat—and can now call tools, access data, and act across production systems—we must ask: are we safe? As agents gain access to production systems, identity, policy, approval, and evidence become essential safeguards around every action they take. We are already seeing the consequences of ungoverned access: production databases wiped, unauthorised MCP servers accessed, and privileged integrations expanding an organisation’s attack surface. Speakeasy’s AI Control Plane helps organisations govern AI access to production systems, with identity, risk, and observability at its core. This talk explores the problems companies face as agents gain real authority—and how to build controls that limit the downsides. The goal is to make agent authority explicit, scoped, and provable, so organisations can give agents permission to act.