Blog
Short notes on what I deal with at work: AI, architecture, agents, adoption into real processes.
Agent Capability Supply Chain
Agentic systems gain a new architectural layer: capability becomes a versioned, verifiable, governed artifact — almost like a software component.
Agent systems are moving from loop-centric to contract-governed architecture
The first AI agents were built around a simple loop. Once an agent runs long, changes system state and spends resources, a managed architecture grows around the model: intent, capabilities, policy, evidence, validation.
It's not the model that decides, but the environment around it
A year ago, an agent was assembled from a model and a list of tools. Now a separate layer has grown up around the model, and it is this layer that determines whether the system works or falls apart on the third step.
AI in business processes: where it pays, where it's hype
Documents, support, monitoring: AI pays back fast there. Full autonomy without control: that is where the losses begin.
Skills, agents and the future of engineering
Expertise gets packaged into skills, routine goes to agents. What is left for the engineer, and what is worth learning right now.
ADR and SDD: documents that don't gather dust
Short decision records and specs before code. Why they matter even more in the age of agents, and why fat specifications are dead.
Where AI actually helps in medicine
No fantasies about replacing doctors: routine work, transcription, a second pair of eyes on scans, and patient follow-up. And why responsibility stays with the human.
AI on the factory floor: don't start with the model
How to bring AI into production cycles without pain: process and data first, a pilot on a boring bottleneck, and only then model debates.
One agent is fine, an orchestra is better
Why a zoo of ten agents never works, and what actually fixes it: an orchestrator, context boundaries and quality control.