Building The Chief of Staff Agent 80x Faster

Host Intro (Gene Kim)

All right. To introduce the next speaker, I've got to tell you about how I spent a couple of days in Munich earlier this year, where I got to hang out with a whole bunch of super switched-on people. And I'm going to show a slide. Imagine a picture of a whole bunch of people hanging out in Munich, where I got to meet just a whole bunch of executives who are using vibe coding.

It was an event that Philipp Göllner, the CEO and founder of XALT, set up. But the coolest part was the weekend before that, where Philipp decided that he was going to set me up my own Open Claude on a Mac Mini. And I was like, "Oh my gosh, this is the best ever." This must've been June. I had been watching all the Open Claude buzz on Twitter, but it just seemed too intimidating to set up. And holy cow, it was so fun. But that was the right call — it was kind of rough sledding, and might be true now, might not be, but it really helped to have someone who's done it before do the Open Claude install.

But that was the birth of Marvin, my Open Claude. And it opened my eyes in terms of what's really possible when you have a persistent thing that's always running, has access to a browser. And it actually allowed me to start thinking about solving some problems that were actually becoming urgent — that I'm not going to go into much explanation about, but it does involve LinkedIn.

But what was really surprising was that my boss and wife, Marguerite, also loved using Marvin Open Claude. And there was something that happened that resulted in a bit of an emergency for 36 hours, where I realized that I forgot to file for a visa to go to Brazil. And it became a full-blown emergency that started around 2:00 PM. And at 2:00 AM, Marguerite says, "All right. We're really in trouble. You should fly down to San Francisco, go to the Brazilian consulate, and see if there's an expedited path to get to São Paulo." And so suddenly, we need so much thinking going on — and enter Marvin, who is a combination expert in visas. They know what consulates are for and what the difference between a consulate and an embassy. And could also help plan, okay, what if we actually got a visa? How do we get from San Francisco to São Paulo? Which was not definitely part of the plan.

Anyway, so just watching Marguerite interact with Marvin on the side, and interacting with this during this emergency, was really mind-expanding.

So I will leave you with one piece of advice. If someone like Philipp offers you a Mac Mini, the answer is not the one that I gave him, which is, "No, thank you. You're too kind." The only correct answer is, "Yes, thank you. I'll take it." Because I learned that it still takes six months to get a Mac Mini.

So anyway, I say this all because one of Philipp's ideas — he thinks so deeply about what it means to be AI-native, how to do really audacious things, including getting all of you an Open Claude set up for this conference. So here to tell the story of how that Open Claude Chief of Staff agent came to be, is Philipp Göllner.

---

Philipp Göllner

Hi, everyone. So yeah, it was so much fun to have Gene in Munich and work with him on the Open Claude. And what we have done together there is building the first XALT AI Cycle — and that's a meeting in Munich where we have like-minded leaders like you guys to share their advantages with the craziness of AI these days.

So what we saw is one app from Robin. He's a CEO of a small airline, very small company, chartering private jets in Germany. And he said, "Oh, I need a new brand." And he basically built, with his Mac Mini, on the first day, a new brand experience. Then he thought, "If I have a new brand already, why shouldn't I have a better app — like American Airlines or like Lufthansa — to have my clients use an airline app?" So he built this app on his Open Claude.

And I think the key thing you see is real people — like CEOs — can build even more with Open Claude, prototype even more than before. It might not be ready for production, but prototype even more than before.

So we have here one example which we saw there from Leonard, who is here right now, and worked on his time-booking app as a marketing manager — no idea about coding, but he worked on an app. And this is what we saw in this room with Gene and all the leaders, talking about how this is all going, how we can survive this era, right?

So out of this crazy idea in Munich came the idea to build a Chief of Staff for everybody here for the conference. And this is the story of how we built the Chief of Staff and how we became 80 times faster in software development on this project. So it's a high-risk, high-reward bet on AI, basically.

So we started like every team — a little copilot, six people. Jürgen sitting here, Ben sitting here, Mark, Mike, and Paul, and two working students. And we were trying to have six full-time equivalents to quickly build this Gene experience in six weeks, basically.

But then something happened where we stopped the two-to-three-times-faster way of working, and we figured out, "Oh my God, wouldn't it be cool — couldn't we do this feature?" And Gene and me were thinking: we need matchmaking, we need questions for the room, we need the program in the app, we need a briefing screen to show all this stuff, and we need an IT Revolution brain covering all the talks and all the books of IT Revolution in this thing. So it answers with quotes from Gene, right? So if you try it, ask it for Gene's quotes — it knows. And yeah.

Three weeks before the event, the team got really scared. So what happened to the backlog? You all know these little burndown charts, right? Where you have the timeline — that's the gray line on the bottom here. And I think the feelings of the team are not overpromised if I say: how do we keep the quality? It's too many features and a limited time. How do we even finish? And, "I want to get out of this project — this is crazy, this will never work," right?

But what did we have? Yeah, the team said what you and Gene want is impossible, basically. But we thought: don't fear ambitious goals, right? That is what we learned in March in San Jose. Use the FAFO thing. And we said, "Okay, might work. We take it on. And if we fail, we fail big. Let's give AI all the power."

So what we had — what Jürgen built in the first three weeks of the project — he built, based on our skills of running enterprise IT, a really rock-solid DevOps process based in the project. So we followed everything which is written in The DevOps Handbook, right? We implemented 6,000 test cases. We had one Jira ticket, one code change. We had every PR reviewed. We had infrastructure as code, so everything could be rebuilt any time. All the goodness of The DevOps Handbook, basically.

But still, we figured out we need a 100-person team, and we only have a five-person team. So what we did is we switched the power over to AI. We switched away from everybody having this copilot, and gave it to an orchestrator. We told Claude, "Take this project, take the Jira tickets, do whatever you want, and don't interrupt us. Create pull requests and push it out to production. It will hopefully work."

So one weekend later, this is what happened to our burndown chart. We basically were back down to scope again. No problem. We have everything under control. Agents write all the code — and then it was shockingly stable. We thought Opus 4.5 with Codex reviews did all the magic, basically.

But you might ask, how is this experience relevant to the enterprise? So we built for Bosch — the big company in Germany building all the kitchenware — we built an IDP platform and worked on a developer experience. And we made a developer self-service on top of the AI to really have the developer have a good experience. And all this was implemented with The DevOps Handbook again. And you won't believe it — DevOps is still relevant. AI also likes good DevOps practices, because it's actually developers.

So we were thinking, is it really time to write the next IT Revolution book? When you search now for developer experience platforms, you get a lot of advertisement, but nobody says "platform engineering for AI agents." So suggestion to Gene: next book — it doesn't exist yet.

---

Chief of Staff, when it hit the wall — obviously, it's not all great. So we found one change in this weekend's batch which completely destroyed everything. We had 50 servers deployed that never showed up in our back end. So we had Chief of Staffs doing nothing — not connected, not recoverable. And we had Ben here, who spent five days, 14 hours a day, sweating and saying, "Oh my God, can we recover this project for Gene?" So if you think about it — give it all to the AI agents — some changes still need a person, I will say.

So after fixing that, we saw a continuous movement in the app: 20 pieces of feedback a day into the AI agents, and 20 features out every day.

And we hit another problem. Ben was still reviewing the security fixes and the hard parts of the tool. But we hit a really different problem of who is actually available via billing. And I still found some features today which I hadn't seen before.

So Helion came up with the idea here, joining two weeks ago after studying and directly dipping into this project. And he found out: wouldn't it be nice to have a little YouTube channel of the changes the agents have done in the app? So we built a little tool to go through all the tickets as it's building them, record a before-and-after video, and channel that out in the morning to the team for quick feature review — where you can really review 10-second before-and-after videos of all features. And yeah, that was really helping, because nobody reads 5,000 lines of code in a PR and figures out what it does. But a little video of before and after was really helpful.

So how did this go over the next days? We had this idea — this feature in the app — and Gene reported feedback with the "I'm Lost" button in the app. It was 9:28, the feature went in. 9:38, there was a ticket and everything. There was a pull request, and there was the code. And 10:13, it was on staging and the video was ready. And whenever somebody approved the video, it went out into production.

So one feature which we see here is the edit button in your profile view when you check it out. There was an edit button — you needed to hit it, very small and very hard to find on mobile. So we got an after video showing the button disappeared, and you can now just type into the text, and every change of the field saves the data. Very nice feature — and a very low-risk change also; it didn't destroy the app. And so somebody approved under the video, "approve it to production," and it went out into production.

And with this speed improvement, we really made it so that we could see — I would say — 80% of all features land and push out to production very fast.

So the learning from that: split the backlog or the tickets by risk. And I heard a lot of good talks in the last two days about how to judge this risk, and push the high-risk things through a Ben review and the low-risk things through a movie review, basically. And when you look at this from an enterprise side again — in this example — we work for the biggest bank for the online banking login in Germany and made this stable and releasing faster. So we are trying to release always faster. And what we learned from this year is we believe even in the biggest regulated industries, you can push out small-risk changes maybe the same way, if you use investments in documenting everything in the development process automatically and really have it ready for review.

---

So how did the math behind the 80x come about? You surely want to know how six people achieve 80x. So we were calculating: four people — one person giving feedback, one customer, and four developers working on the features. We said what the AI can do in one day, it takes us 20 days — one sprint, one four-week sprint, basically. And that calculation made us believe it's really 80 times faster, and that came from the developers. So four weeks is now two hours, in the end.

So why is this North Star prototype so important for this conference, and for you to try it, and for the whole industry? These agents give us CEOs even more power of having another wrapper around our AI, which helps us ask — and as long as it thinks it helps us and gets us to the right level. And we believe this will be the next wave in AI, basically.

So most companies are still on wave one of personal productivity. Then some companies are already running AI in their processes and their apps and their business software. And I think the next AI phase is the enterprise brain and the agents, which have all the context and can guess even more what to enter into the prompt — and get you away from the Claude code console again and back into the meetings.

So in the enterprise, we have done something similar to this Chief of Staff. We have done a project portfolio management agent which knows everything about the project, what is going on in Jira, and grows with you, understands how the team works, and helps you with not failing in projects. Because every project manager says, "An earlier warning as a manager or a project manager would have helped me to fix the project still." And if you have your agent sitting on top of the project, you don't have to micromanage people, but you can see anomalies — the team got pulled away into another project, the team got pulled out of your project. And this is another use case for an agent like that, which we call the early warning system for managers. The pulse of your project — it notices changes.

---

What is next after the 80x now? So we found one feature here at the conference. Gene said the agent conference message is a very big AI blob — it's not helpful, it's not useful. So I thought, "Oh my God, the Wi-Fi is really not allowing SSH." But I told Claude, "Run a self-learning loop on the discussion of the main people at the conference, compare it with the agent conference message, and learn what you're doing wrong to become more human." And I think it's really what Steve said yesterday — measuring how humans work and making the AI understand what we want from it is really helping. And now when you look in the agent conference channel now in your Slack, you will see it's already way better and way more human than yesterday.

So this is probably the next thing: that we let agents run on top of the software, and let them learn the low-risk features, let them improve the low-risk features for humans.

So if you want to try the agent conference more and go into your Chief of Staff and play around with it — really, really important — because this is the demo case for our meeting in November with Peter Steinberger. So as the second speaker after Gene, we got Peter to come to Munich in our Claude AI Cycle, and he will talk about how the Chief of Staff ran the conference with us. So please try it more. You have four weeks still of access to the experience. Get out your learnings, put in your screenshots, and try it out even more.

And before I forget — what I am missing in my brain right now — this is the IT Revolution brain — I want to know: what audit steps did I forget? What audit steps do you have in your pipelines right now which we could train the AI agents to know? And how do you judge high-risk tickets? Because our system didn't catch the one Ben needed to fix. And what do you trust AI to take over completely, and what do you still want to review? That is really important to me.

And if you want to talk about this further, ask your Chief of Staff to reach out to me. He will introduce you to me.

Thank you very much.