Engineering writing from inside the build.
Reference docs on how Svolta builds and runs AI agents for real work. Engineering patterns, not opinions. The quality checks, model routing, and operating discipline behind the systems we build.
The durable patterns.
How we run quality checks, how we route between models, and how we operate agents once they are live. The insights archive carries field notes and opinions. This is the reference.
- 01
How we run evals in production
The four-layer eval suite that gates every Svolta agent release. Golden datasets curated with subject matter experts, regression and behaviour tests, drift detection, and the reasons LLM-as-judge alone never ships the call.
Mac SweenyFounder12 May 2026 - 02
How we route between models
Outside of evals, no infrastructure decision moves cost and quality further than routing. How Svolta routes across Anthropic, OpenAI, and small local models by cost, latency, and accuracy, with deterministic fallbacks and hard cost ceilings.
Mac SweenyFounder22 Apr 2026 - 04
Our observability stack for agents
What ops leaders actually look at on Monday morning. Per-call traces, live eval scores, drift alerting, cost dashboards. The day-one surface that lets a senior team run a Svolta agent without paging the build team.
Mac SweenyFounder15 June 2026
Want this kind of system in your stack?
The shortest path from reading the architecture to a costed plan for your own. Bring one workflow. A free 30-minute Consultation to scope the path, an Audit to map it.