AI-Native SDLC: Code Accelerated, Delivery Didn't
Anthropic's playbook—and a transfer test for regulated companies
Slide contents
1. AI-Native SDLC: Code Accelerated, Delivery Didn't
Anthropic's playbook—and a transfer test for regulated companies
2. Two people, one transfer test
Alexander Polomodov × Anton Kosterin, T-Bank
3. Playbook, not comparative research
A useful operating model, not a universal result
4. The queue moved beyond coding
Planning, review, testing, and release absorb the load
5. The old relay hides waiting
Documents decay between functions
6. Replace handoffs with a loop
Every stage leaves a verifiable artifact
7. Six stages, one operating loop
Not stage automation, but decision continuity
8. Artifacts become the audit trail
intent.md → spec.md → plan.md → diff/tests → review → incident
9. Key decisions stay with people
Product, architecture, code ownership, and production
10. Start from intent, not a ticket
Problem, boundaries, outcome, and decision owner
11. Policy enters before code
Architecture and security constraints become actionable
12. Plan mode shrinks blind work
Files, dependencies, steps, and checks are visible first
13. Team memory belongs near code
Commands, conventions, and verification paths are versioned
14. Advice suggests; hooks enforce
Deterministic checks remain deterministic
15. Review capacity caps parallelism
A worktree isolates code, not attention
16. Testing moves into the loop
Small diff → fast check → correction
17. Incidents become regression cases
The task set evolves with the system
18. Review goes both directions
Agent finds defects; owner accepts risk
19. Agent stops at production gate
Sandbox, scoped credentials, branch protection, rollback
20. Operations feeds the next intent
Monitoring is deterministic; changes flow through PRs
21. Choose one source of truth
Other systems retain a link and status
22. Measure flow and control together
Baseline, bottleneck, outcome, risk, and learning
23. Start with one bottleneck
Baseline → pilot → gates → evidence → scale