How We Built AI Agents for Our Product Organization (and How You Can Too)

Most talks about AI agents assume you're already working with Claude, custom agent frameworks, or some other cutting-edge setup. But a lot of organizations don't have that infrastructure yet, and there's still a lot of power sitting in tools you probably already have access to, like ChatGPT, Copilot, or Langdock. You don't need a fancy setup to get real value from agents.

Over the past months, we've set up a number of AI agents in Langdock to support our product organization, from repetitive research tasks to recurring decision support. This talk walks through that setup: which agents we built, why, and what we learned along the way.

It's designed for people who have never built an agent before. If you've used ChatGPT or Copilot with a well-crafted prompt, you're closer than you think. The concepts are largely the same, and I'll point out where Copilot and ChatGPT work similarly so you can apply this regardless of which tool your organization uses.

I'll also cover a problem that doesn't get talked about enough: agents don't stay reliable on their own. I'll show why you need someone responsible for each agent, checking that its output isn't quietly drifting, and testing it regularly against example data, much like you'd maintain any other piece of critical infrastructure.

This is a hands-on talk: I'll build and adjust an agent live in Langdock, so you leave with a concrete setup you can bring back to your own organization.