Don't Make Your Agent Guess — Impressions from AI Engineering London

Last week I had the privilege to talk at the AI Engineer conference in London. It's amazing to see a new community emerge and be part of it. I had lots of great conversations with the audience and many amazing speakers. We were also proud supporters of the conference, which made it even more special to contribute and meet so many people shaping this space. I really appreciated the way they are celebrating AI innovation from Europe. swyx and the whole AI Engineering team are on to something here.

While there's a lot to unpack, I'd like to share some of my key takeaways while they're fresh:

1: Agents Need More Than Chat

One of the last talks was by Jacob Lauritzen, CTO of Legora. He shared some of their lessons learned when building vertical AI applications: work is changing. While it used to be mainly about doing the work, we are getting into a state where it's more about planning and reviewing.

Jacob talked about verifying work and his examples of decomposing tasks triggered a lot of ideas for me. He also explained high-bandwidth artifacts. Tabular view for the win. To all the business folks: your spreadsheets are going nowhere. ;-)

See his full talk here: AI Engineer World’s Fair / Jacob Lauritzen

2: Agent Archetypes

Malte Ubl, CTO of Vercel, kicked off the conference with a great intro on how we are moving into a new software paradigm: from sites to apps to agents. And that agents make many more automation use-cases viable, so we will see a lot more software.

What I appreciated most were his four agent archetypes: 1: Always Running, 2: Compress Research, 3: Surfaces Hidden Information & 4: The Boring Things.

See his talk here: AI Engineer World’s Fair / Malte Ubl

3: Don't Make Your Agent Guess

There is a lot for us to discover on how to build agentic systems or as Mario says: "We are in the fuck around and find out phase". But when I was rehearsing my talk with Ivan Pedrazas, we realized that when architecting our systems we tried to make it easy for the agent.

This starts with a good prompt, but touches other parts of the harness. Where we recently saw the biggest impact was in tool design: exposing the right functions, writing good return statements especially in error cases, and reducing the number of things the agent needs to figure out on its own. Ivan has some great stuff cooking there. Stay tuned.

In my talk I covered PI and how we are using its malleable properties when building agentic systems like we have recently done for a client with our sales quote generator.

If you are interested in these topics, don't hesitate to reach out. And check out the full talk here: My full talk

And finally, don't miss Mario Zechner's talk. A great dive into coding agent harnesses and how to build open source software in these crazy times: Mario's talk

BTW: This was our first officially supported conference. We couldn’t be more proud. Many more to come.