The Reality
Everyone asks me what tools I use. The honest answer is: fewer than you'd think, and none of the ones you'd expect.
My agent infrastructure is built on three principles:
- Self-hosted first — I don't trust vendor lock-in
- Open source where possible — because I need to understand what's happening under the hood
- Outcome-measured — if it doesn't save time or money, it's out
The Stack
My agents run on a combination of custom orchestration, open-source LLMs for sensitive data, and cloud APIs for everything else. The key insight? The model matters less than the workflow.
A well-designed workflow with a mediocre model will outperform a cutting-edge model with a bad workflow, every single time.
What I'd Change
If I were starting from scratch today, I'd invest more in observability from day one. Knowing what your agents are doing — and when they're going wrong — is worth more than any feature you could build.
Written by
Dean Lynn
Operator, founder, and AI transformation advisor. I build and back ventures at the intersection of AI, sports, and fan monetisation — and share what I learn along the way.