I was offline. Here's what happened when I came back.
This edition covers the AI engineering conference I attended, my talk on multi-agent orchestration patterns, a new article on context graphs, and a deep-dive video on AI latency
Hello everyone,
I owe you an explanation for going quiet last Saturday.
We took an Easter break as a family - properly offline, no laptop, in the English countryside. It was superb, the rolling green fields actually delivered on the promise. - and it was sunny ☀️
Right. I’m back. And there’s quite a lot to catch you up on.
I was at an AI engineering conference this week
And I gave a talk, it will be out soon. This is the first time it happened in Europe and I got to meet so many smart, talented founders and engineers. It was awesome experience. In-person events are irreplaceable.
I attended some fabulouse sessions on cutting-edge stuff on AI. And of course OpenClaw dominated the discussion.
Check out the conference here: https://www.ai.engineer/europe
I also made it to the online track of the conference
And the topic was something I’ve been working towards for a while - Multi-Agent Orchestration Patterns for Production.
The core argument: the field is moving fast, but most teams hit the same wall. They build multi-agent systems like they built single-agent systems. Same assumptions, same trust in the “it works in the demo” signal. And then production arrives, and nothing holds.
The talk walked through choreography vs orchestration, immutable state patterns, circuit breakers, and why distributed systems thinking is no longer optional if you’re building agents at any meaningful scale.
Check it out.
A piece I wrote just went live
This one has been in the works for a while, and I’m genuinely proud of it.
I wrote a guest article for Context & Chaos introducing two concepts I’ve been developing from my work with regulated enterprises: context drift and the evaluation graph.
When an AI system gives you an answer, that answer wasn’t produced in a vacuum. It was produced against a specific version of your world - a specific definition of what “active customer” meant that week, a specific policy that was in force that month, a specific dataset that may or may not still exist.
Think of it like this: imagine a doctor’s notes. It’s not enough to record what prescription they wrote. You also need to know what guidelines were current that day, what the patient’s history showed at that point, what the lab results said. Without that context, the notes are incomplete. Enterprise AI has the same problem. We’re recording the prescription. We’re not recording everything else that informed it.
This is original IP, and I think it’s going to become a recurring theme in how regulated industries think about AI governance.
Here is the article:
New YouTube video: The AI Latency Stack
While you’re in content-consumption mode this weekend, I also want to point you to a video I put out recently on AI application latency - it keeps coming up in conversations and I wanted to have something concrete to point people to.
The short version: after your AI system ships to production, the model is almost never the problem. It’s a set of architectural decisions - streaming, database writes on the critical path, cold starts, context window bloat, prompt caching, sequential calls that should be parallel - that compound into something that makes users give up and go back to the manual process. The video walks through seven of these layers and how to address them, without swapping models or changing vendors.
Worth a watch if you’re anywhere near a production AI deployment right now.
That’s it for this week.
A lot happened in a short space of time, and I wanted to share it with you directly. As always - reply if anything resonates, or if you’re wrestling with something I touched on.
Talk soon,
Sandi
P.S. If you’re new here - welcome 🎉. AgentBuild is a community of practitioners working through the real challenges of getting AI into production inside large organisations. Every week I share practical, grounded thinking from the people doing this work at the sharp end. The goal is never theory - it’s always: what can you use Monday morning.
Ask your friends to join.
More valuable content coming your way.
Thanks for reading agentbuild.ai! Subscribe for free to receive new posts and support my work.