Research Insights Made Simple Episode #31 Materials: AI4SDLC — What I Would Do Differently (Series #AI4SDLC)

I’ve collected the materials from my solo Research Insights Made Simple episode of September 22, 2026. It is the director’s cut of my Deep Tech Night lightning talk: 15 minutes at the conference became almost an hour and a half on air.

A project management assistant works well on its own team’s tasks but gets lost on projects spread across several trackers. Some users simply stop opening it without saying a word. Should you pick a stronger model? First you need a set of real tasks on which such a switch can be tested.

In the episode, I covered the following points 1️⃣ The ownership boundary and build/buy decisions Rent models and the basic agent loop, adapt context and adapters to internal systems, and keep permissions, agent identity, and quality evidence under your own control. 2️⃣ Permissions and tools An agent’s ability to plan a fix does not give it the right to carry it out, and mechanically wrapping OpenAPI in MCP does not make a tool convenient for an agent. 3️⃣ Quality checks Demos, traces, and eval sets answer different questions. Successful and failed tasks are worth keeping and rerunning every time the model, context, or harness changes. 4️⃣ Economics Once work is broken down into verifiable engineering episodes, you can try smaller models and separate planning from execution.

And custom additions will have to be rechecked regularly: the next model may handle the same task without the accumulated workarounds.

Episode materials: 📌 Episode page with timestamps 📊 Talk slides 🎬 YouTube, VK Video 🎧 Podster, Yandex Music, Apple Podcasts 📝 Written recap

Which part of your AI stack have you already made your own? And when did you last check that it still beats the ready-made option?

#AI4SDLC #AI #Agents #Architecture #PlatformEngineering #Evals