Research Insights #43 Materials: Agent Skills and Open Knowledge Format

The materials for Research Insights Made Simple #43, from October 22, 2026, are ready. Maxim Smirnov, author of the Telegram channel “IT Solution Architecture” https://t.me/it_arch, joined me to continue the conversation about Loop Engineering. Writing a SKILL.md is easy. The more interesting questions are when an agent will choose it, whether the instruction helps, and which knowledge it relies on.

What we discussed

🔸 Why a skill catalog alone doesn't solve the problem
A broad description can lead an agent to choose a skill where it doesn't fit. A useful, concrete procedure often depends on the team's rules and context, so moving it elsewhere takes some adaptation.

🔸 How to measure whether a skill helps
Collect tasks where you had to fix an agent's work by hand, then compare the old skill, the revised skill, and the agent working without it. Look at quality, trajectories, time, and token usage. Repeat the check after switching models.

🔸 What knowledge several agents need
A technology radar, architectural decisions, and how the system is put together. We looked at LLM Wiki and OKF: a connected representation of a domain that agents can draw on across tasks.

🔸 Where document work remains
A small knowledge base and search across an organization's documents pose different challenges for updates, relevance, and access. We aren't choosing a universal replacement for RAG.

🔸 Why a source link is only the start of trust
It matters how a derivative was created and what was checked. We discuss data provenance and proof of computation in the OKF specification. The episode doesn't demonstrate a working verification mechanism.

Another thread is the maintenance all of this needs. Instructions and knowledge both go stale. Even a regular agent walkthrough of a knowledge base needs quality criteria and feedback.

Episode materials
📌 Episode page with timestamps and episode slides
🎬 Video: YouTube and VK Video
🎧 Audio: Podster and Apple Podcasts
📖 Written summary
📖 Specifications: Agent Skills and OKF

Which skills have already survived a model switch and remained useful for you? I'd like to hear about successes and about the ones you had to discard.

#ResearchInsights #AI4SDLC #Agents #Skills #KnowledgeManagement #Evals