![A simplified trace of one WhatsApp turn. A chat bubble reads “paid the piano lesson for [NAME_1]”, with the name replaced by a placeholder. Two chips read $0.036 and 7 calls. Below, a waterfall: one long bar for the turn, then alternating blue model calls and short orange tool calls.](/_astro/obs-article-cover.Bny3A09H_Z1Wb1gF.webp)
Watching an Agent Work: Observability and Evals for Lukre
How I added LLM observability and an evaluation lab to Lukre, exploring tracing, privacy, model costs and the limits of early evaluation results.
Writing / Category
Building AI systems that work beyond the demo. Agents, retrieval, evaluation, and the engineering in between.

Browse by category, or page through the archive below.
![A simplified trace of one WhatsApp turn. A chat bubble reads “paid the piano lesson for [NAME_1]”, with the name replaced by a placeholder. Two chips read $0.036 and 7 calls. Below, a waterfall: one long bar for the turn, then alternating blue model calls and short orange tool calls.](/_astro/obs-article-cover.Bny3A09H_Z1Wb1gF.webp)
How I added LLM observability and an evaluation lab to Lukre, exploring tracing, privacy, model costs and the limits of early evaluation results.

From a WhatsApp message to a financial record: the boundaries, contracts and application logic behind Lukre.
Adding agents moves the hard part to coordination: who decides the work, who keeps state, and who checks that it is done.
A 2024 reflection on AI, programming, and the importance of collaboration: code got cheaper, and design, communication and business understanding matter more.