Episode materials with Anton Kosterin: how do you close the AI-native SDLC loop? (Rubric #AI4SDLC)

The materials for episode 29 of Research Insights Made Simple, held on August 27, are ready. Together with Anton Kosterin, a principal engineer at T-Bank’s R&D center, we examined Anthropic’s “The AI-Native SDLC Playbook”: what has to change around coding before agent speed can become delivery speed.

We discussed:

• Why the playbook’s value lies not in fundamentally new practices, but in assembling them into a sequence with examples and templates — and why a vendor guide still needs to be tested in an organization’s own context;
• How intent.md, spec.md, and a human-approved plan.md turn an idea into a versioned artifact chain that also serves as an audit trail;
• Why a team needs CLAUDE.md, skills, commands, hooks, and specialized agents — and what upfront investment and organizational trust such a system requires;
• Why the constraint moves to problem framing, review, testing, or release once Build accelerates, while deterministic CI/CD checks and continuous evals should independently catch violations and recurring failures;
• Where the automation boundary lies: hooks enforce verifiable prohibitions, while review, engineering judgment, and acceptance of residual risk remain human responsibilities;
• How Maintain closes the loop: an incident can produce a new intent, while a runbook can become policy-bounded agentic self-healing; the economics are better measured through accepted tasks and avoided rework than through tokens alone.

All episode materials:

• Episode page and slides
• Video: YouTube, VK Video
• Audio: Podster, Yandex Music, Apple Podcasts
• Text: concise transcript

Thank you to Anton for the conversation. If the episode leaves you with questions, post them in the comments; I will collect them for a follow-up on bringing AI-native practices into large engineering organizations.

#AI4SDLC #AI #Agents #Engineering #Architecture #DevOps #Research