Changelog

In Progress
Feb 2026

DSPy Optimization

Applying DSPy MIPROv2 optimizer to our failure detection pipeline, discovering that switching to DSPy native format alone surfaced more failures than our carefully authored instructions ever did.

Vladimir VučkovićVladimir Vučković
Shipped
Feb 2026

Boring SaaS Work

Implemented Auth & Billing into our observatory platform for operating moyai agent reliability solution standard SaaS functions.

Robert HommesRobert Hommes
Shipped
Feb 2026

Unleashing the Grep Tool

Dual-mode evaluating agent now runs on a single grep tool with zero prescriptive guidance. Opus searches the traces blindly and discoveres new anomalies entirely on its own.

Vladimir VučkovićVladimir Vučković
Shipped
Feb 2026

Implemented PlanetScale

The world’s fastest and most scalable cloud database is PlanetScale. Our AWS estate is connected to our managed PlanetScale database using AWS PrivateLink for maximum security.

Robert HommesRobert Hommes
Shipped
Feb 2026

Self-Evaluating Loop

Added a new agent to our anomaly detection workflow that reduces bias and increases variance in classifications. A "blind" pydantic-ai ReACT loop on Opus 4.6 reanalyzes logs without seeing prior results, then compares both assessments and emits the verdict - ensuring completeness across the pipeline.

Vladimir VučkovićVladimir Vučković
Shipped
February 2026

Adaptive Log Exploration

Built an adaptive token sampling strategy for large agent traces, prioritizing records from the beginning, middle, and end of context windows to preserve signal while fitting token budgets.

Vladimir VučkovićVladimir Vučković

Community engagement

We build in public. Want to request something or upvote a feature? Join the conversation on: