The hidden cost of retries in agentic AI systems
How provider, application, tool and workflow retries combine to increase both latency and cost — and how to put a budget on them.
Blog
Notes on AI in production, distributed systems and reliability.
Written by Barnabás Szőke from production work. Experiments, failure analysis and methodology — no listicles.
Stop measuring AI cost per token
Why cost per successful workflow is a more useful production metric, and what it exposes that provider invoices hide.
How to decide whether a cheaper model can replace your current model
A practical process using representative workloads, evaluation thresholds, latency and economics — with the gates that should block a release.
Running one of these problems in production?
Retry behaviour, inference economics and model evaluation are part of the AI Systems Audit.