AI-Native SDLC: Code Accelerated, Delivery Didn't

Alexander Polomodov and Anton Kosterin examine Anthropic’s The AI-Native SDLC Playbook. It is an operating guide drawn from the vendor’s own practice rather than a comparative study. Much of the material will be familiar to experienced engineers, yet the ordered workflow, concrete examples, and reusable templates make it a useful starting point for teams without an established agent process.

The playbook connects six development stages into one loop. Planning turns a product idea into intent.md; design produces spec.md and incorporates organizational constraints; build starts from a plan.md that a person has actually reviewed, then relies on CLAUDE.md, commands, hooks, skills, and specialized agents. Faster code generation does not remove the constraint—it shifts waiting into framing, verification, review, or release.

Early tasks pay for constructing the harness, while team trust and approval from control functions require evidence. Deterministic CI/CD gates must enforce security, legal, design, and other requirements independently of the model. Continuous evaluations protect the agent loop when its model or harness changes, and hooks make non-negotiable rules executable; neither mechanism removes the human reviewer’s accountability for residual risk.

Maintain is the playbook’s most distinctive stage: an operational incident can create fresh intent and enter the same development loop, extending a runbook toward policy-bounded agentic self-healing. The useful economic unit is an accepted task and the rework avoided, not token spend alone. A practical adoption path is to start at one observed bottleneck, borrow the relevant template, validate it in local conditions, and keep production authority with accountable people.