Starting here

Most of what I've learned this year came from systems that didn't behave the way they did on my laptop once they were running inside a bank. The gap between a demo and a deployment is where the interesting engineering is, and almost none of it gets written down.

So this is where I'll write it down. Short notes, mostly: a thing that broke in production, how the traces and evals pointed to the cause, what changed in the harness afterwards. Client names stay out. The mechanisms are what matter.

Three threads I expect to keep coming back to:

• Evaluation as a release gate. How to turn "it seems fine" into a regression check that can actually block a deploy, and what that costs in latency and maintenance.
• Harness design. Orchestration loops, tool interfaces, context assembly, and the deterministic fallback paths that make an agent's output auditable to a risk team.
• On-prem reality. What integrating with a legacy core banking system does to your architecture, your model choices, and your sleep.

If any of this is useful to you, or you think I've got something wrong, email me.