The Flywheel

Sixteen AI agents built a working C compiler in two weeks. $20,000. Sounds like another day in the AI hype cycle. But the most important thing they needed? Something as far from hype as possible.

Week 5 of life post-MAS. (Links to week 1, week 2, and week 3, and week 4).

Fewer conversations. More doing. Teaching, webinars, prep calls, negotiating contracts. Tedious stuff.

At the end of a week, I read this article by Anthropic on “agent teams”. Which they used to describe multiple Claude instances working in parallel on a shared codebase without active human intervention. Somehow this crystallised the week for me.

Let me explain why, by stepping through my week.

The familiar. An ex-boss from MAS, catching up. Being a busybody at a meeting between folks developing AI for the financial industry and a Professor proposing a foundational model. Folks I am working with from an institute in the UK, discussing an AI risk management training module. An ex-colleague now doing her own thing - compliance consulting, wellness coaching.

New connections. An anthropologist studying tech, finance, and money. A sustainability leader thinking about where AI fits in his world. A founder still shaping his plans. A trading community, sixty strong on a webinar where I was on the panel, asking hard questions about AI in markets.

The throughline - flywheels.

Agents in Trading

Friday evening. Joined a webinar on AI in trading as a panelist. Over sixty participants. The questions were interesting. Someone asked about trading venues made up entirely of agents. Many of the other questions were also about the impact of AI agents on trading.

But the most meaningful points weren’t about the frontier. It was about the boring fundamentals, continuing to be important even as we shift to this new world. Systems that failed not because the AI decided badly, but because the basics were missed. I shared my boring perspective. That while the technology changes, the sources of failures don’t.

The webinar was organised by the FIX trading community - an association that oversees the FIX messaging protocol for trading, which includes messages for algo certification and testing.

If you think about it, such boring infrastructure is a flywheel. Once set up properly, the flywheel makes things happen. Perhaps more effective than any regulation. Even ones I wrote.

Such flywheels could perhaps also be the way to go for AI testing.

A Different Lens

The most unexpected conversation this week was with an anthropologist.

She studies how technology and finance reshape people’s relationship with money. Not the systems. Not the models. What the technology does to the people interacting with it. If you think about it, money is AN interaction interface. Its form shapes how people transact.

She spent ninety minutes asking me questions. About my experience with money across my career in MAS. From Basel to models to investment risk. About what I thought AI risk management was actually for.

Again, I explained why I think the boring work matters, not because regulators say so, but because without it, systems won’t work.

At some point she asked about my PhD. How I approached the research. I told her about the flywheel I set up. Collected the data early. Set up the experiment. Found the baselines and the state of the art. Those were the bookends. And then I just iterated. Fast.

She didn’t speak of it. But I suspect her work has a flywheel too. Her papers are way more unique than mine. But I imagine each one builds on the last. The fieldwork from one project becomes the foundation for the next question.

The Agent Experiment

Then over the weekend, I read something that tied it together. Anthropic published a piece about tasking sixteen Claude agents to build a C compiler autonomously. But what caught my eye wasn’t the output. It was what the researcher said made it work:

High-quality tests - otherwise the agents solve the wrong problem
Documentation that agents update constantly - so the next agent can orient itself
Change management - new commits can’t break existing code
Monitoring and logging - without feedback from these, the agents lose context and waste time

I read that list and thought: that’s the same list. The same list from the AI risk management guidelines I wrote. The same boring points from the webinar. The same way I approach my AI research. The flywheel.

Interesting. I thought.

Still Finding my New Flywheel

Five weeks out. The arc is settling.

Valleys. Messy middles. Phase shifts. Resonance. And now, the flywheel.

Hopefully I get the flywheel for this new phase of life soon.

What’s your flywheel for life?

#Flywheels #AIRiskManagement #Fundamentals #Transitions #AgenticAI