Multi-agent AI systems rarely work the way most teams expect.
In this episode, self-organizing AI agents build an entire product with no orchestrator in charge, deciding together when the work is done. David Puglielli, senior advisor and software engineer at Flexion, walks hosts Matt Sharp and Holly Fake through an experiment in ensemble coding with AI most teams haven't tried yet. Rather than one agent handling everything, David built four distinct AI agent personas that debate, self-correct, and vote on their own progress, a structure he chose specifically to avoid the risks of a single point of failure.
David explains what happened when he asked multiple agents to each pick a random word, and why they kept landing on the same answer. He also describes a simple AI code verification workflow anyone can try, and the AI test coverage risks worth watching for when agents are left to optimize for a single metric on their own.
This is what AI agents are capable of building when the only thing checking their work is each other, without any single agent having the final say.
06:28 Why David moved away from a single orchestrator
08:38 The random word test that agents kept failing
17:26 Fixing it with seeded randomness
20:00 Turning agents loose on a real build
21:25 Four personas, one team
25:03 The risk of optimizing for one thing
26:34 The driver and verifier pairing
What is The Learning Curve?
The Learning Curve is Flexion’s podcast about navigating change in complex systems — one decision at a time. Each episode explores how organizations can harness technologies like AI not just to follow trends, but to drive measurable impact.
In this season, we’re exploring:
-How real teams are building trust and transparency with AI
-The practical realities of enterprise AI adoption
-Lessons learned from Flexion’s own journey, experimenting with and scaling AI
This isn’t hype. It’s an inside look at the mindset, structure, and strategy needed to lead in an era of continuous change.