GOTO 2025: AI Agents in Practice

Here are the slides for my talk "AI Agents in Practice" at GOTO Conference in Copenhagen. I showed a bunch of concrete examples and demos, and shared some learnings from the past couple of years of agent development.

Abstract:
What is an AI agent, how do you make agents that actually work and do useful stuff? And what are some things to keep mind when doing so - tips, techniques, pitfalls? In this talk Henrik shares experiences from real-life agent development.

This is an updated version of an earlier talk on the same thing, but with some new content, including some crazy demos:

• Made an agent that checked the latest updates in github, posted summary on slack. Then I asked it to call me on my phone (voice call!), describe the changes in Danish, and then pretend to get really angry at me for not speaking danish back. Several people asked me afterwards if that was actually real, because it seem almost unbelievable. Yes it was real, the Voice Call capability in our platform is a combination of Twilio + Open AI realtime API.
• Showed examples of how we've built agents for investment screening, and handling complex scheduling tasks
• Showed an example of an agent building it's own tools on the fly, using Http Request capability (a swiss army knife, essentially a tool to build tools).
  

Always enjoy talking about this stuff! But I have to update my examples every time, because of the crazy pace of development.

For example with yesterday's release of Claude Sonnet 4.5, our agents can suddenly do things that were previously impossible. I was pretty stunned myself, when prepping some of the demos for the talk.

Here are some related links:

• The Agent Design Canvas that I showed.
• Build your own agents on the Abundly platform. Note that the Http Capability is hidden by default, since we haven't yet had time to evaluate the safety implications. But you can ask us to enable it if you want to try it. Same with the Call and SMS capabilities.