Building AI-first: Lessons from Lisbon

Takeaways from our Lisbon AI Week evening with Alpic, Dust, and Speakeasy: the question is no longer whether to use agents, but how much to let them do.

Every AI company is running agents now. The question is how much they trust them to act autonomously.

On September 22, we packed our Lisbon HQ with founders, builders, and engineers to talk about building AI-first as part of Lisbon AI Week. Speakers from Alpic, Dust, Paddle, and Speakeasy showed what their agents can do, and when they report back to their handlers.

The four talks couldn't have looked more different on the surface:

• In "The Owl Doesn't Know Me," Quentin Churet from Alpic demoed Gordito, a language-learning agent built with Alpic's framework for MCP Apps, testing his own Spanish along the way. Its persistent memory knows why he's learning, not just what he got wrong.

• Paddle's own Kianna Love and Jake Duffy presented "Is Your Platform Promptable?" explaining what it took to make Paddle's API-first platform ready for Lovable's AI agent and how this transformed our developer experience.

• In "The Interruption Tax," Nicole Kreider from Dust showed how agents can gather shared context and resolve requests before anyone has to interrupt a colleague, and when they should hand off to a person anyway.

• Tiago Zien-Mendes from Speakeasy walked the audience through real incidents caused by agents' unsolicited actions in "Permission to Act," introducing Speakeasy's AI Control Plane that helps make agents' authority explicit, scoped, and provable.

In essence, they all described the same shift. Agents are no longer just chatbots that answer questions. They're taking real actions in real systems, and every team had its own answer to how far they should go.

Here are four things we took away from the evening.

An agent is only as good as its sources

An agent's mission success depends on complete, accurate context, not on the model it's built on. Give it bad intel, and you either get poor-quality output or land yourself in an endless loop of reprompting and clarifying. Gordito helps Quentin advance his Spanish because it knows the learner and his motivations, tailoring the curriculum to him.

The promptability of Paddle's platform largely depends on the quality of its documentation and API design. Dust's agents only save the team the interruption tax when their shared company knowledge is complete and up to date. Point them at old or incorrect sources, and the human in the loop is taxed twice: once when answering the same questions, and again when untangling incorrect agent responses.

So before you debate which model to use, audit your agent's sources.

Your product is away on assignment

The product you're building no longer lives only on your platform. More and more, it's used through an agent, inside someone else's interface. With Paddle's Lovable integration, developers can build, go live, and manage transactions without ever opening Paddle's dashboard. Gordito doesn't have a website of its own. It lives inside the chat you already use, where the agent decides how it's presented.

In each case, your carefully designed UI with its visual feedback is no longer in the picture. The quality of an agent's interaction with your product now depends on your API, its responses, and its permissions. Speakeasy takes this to its logical conclusion, treating agents as full-fledged actors in a production system.

Now that your product is away on assignment, make sure it can clearly tell an agent what happened and what to do next.

Every agent needs rules of engagement

Just because your agent can do something doesn't mean it should. None of the companies we heard from chose to give their agents full autonomy. When using Paddle with Lovable, an agent can build an entire billing integration, but it defers to a Paddle embed for verification. The agents in Dust's workflows are designed to know when to hand a request to a person, and what to pass on so nobody starts from scratch.

Rules of engagement only work if something enforces them, or else your agents go rogue. Speakeasy's talk showed what that looks like, from wiped production databases to unauthorized MCP servers. Their approach is to give every agent an identity, access scoped to the job it needs to do, and an audit trail of the actions it takes.

So before you send your agent out, decide where its mission ends and how you'll enforce that line.

MCP servers are your agent's gadgets

MCP has settled in as the way agents connect to the outside world. Gordito is built as an MCP App, so it can be called wherever your agent lives, without needing its own website. Paddle runs its own hosted MCP server, so agents in tools like Claude Code or Cursor can investigate failed payments, handle pricing, or analyze revenue. Speakeasy, meanwhile, is focused on keeping the growth of MCP usage in check.

Just as the best APIs set their companies apart, a well-designed MCP server is now a competitive advantage. The details matter: which tools you expose, how clearly you describe them, and what the server says when something fails. That's also what the room wanted to hear more about, with feedback after the evening asking for code-level walkthroughs.

So before you ship an MCP server, make sure it equips your agent clearly, completely, and with nothing left to guess.

The mission continues

Not long ago, the question in a room like this was whether to bring agents into your workflow at all. In Lisbon, nobody was asking that anymore. Every team on stage had already recruited its agents. Deciding what they're cleared to do is the mission everyone's still on.

A huge thank you to Quentin, Nicole, and Tiago, and to Alpic, Dust, and Speakeasy for sharing the stage with us. And thank you to everyone who came along, paid attention, and asked questions. You made our Lisbon AI Week a success.