Anthropic's AI-Native SDLC Replaces Gates With Git Loops

Anthropic published The AI-Native SDLC playbook on August 21, 2026, outlining how enterprise engineering teams can redesign software delivery around agentic tools including Claude Code, Claude Enterprise, and Claude Tag. The central change is organizational: faster code generation moves the bottleneck to planning, design alignment, security review, testing, approval, and incident response.

From staged pipeline to continuous loop

The playbook replaces a mostly linear SDLC with a non-linear loop. Each stage produces a version-controlled, machine-executable, human-readable Markdown artifact, and acceptance of one artifact can trigger the next agentic workflow.

Traditional SDLC	AI-native SDLC
Requirements, tickets, and committee approvals	intent.md synthesized from stakeholder conversations with Claude
Separate product, UX, and systems specifications	Unified spec.md generated against organizational constraints
Manual translation from requirements to code	Claude Code produces an implementation plan.md before editing files
Late QA gates	Continuous evals and generated tests during implementation
Line-by-line review of every pull request	AI first-pass review with human attention on high-risk changes
Manual incident triage	Monitoring agents draft new intent artifacts from production anomalies

Keeping these artifacts beside application code creates a single repository audit trail. Teams can inspect who requested a feature, what the agent produced, which approvals were granted, and how the implementation changed over time.

Six stages and their control points

Plan begins with a stakeholder or originator working with Claude to convert product goals and user friction into an intent.md or proto-spec file, stored in a version-controlled directory such as /intent/.

Design combines requirements definition and architecture in a prompted session. Claude reads intent.md and generates spec.md, while reusable Claude Skills encode brand, security, and UX constraints as hard generation requirements. A product owner signs off before implementation begins.

Build starts with Claude Code in plan mode. The agent reads spec.md, interviews the engineer about the repository, and produces plan.md with proposed file changes and test requirements. Approval authorizes implementation. Repository and directory-level CLAUDE.md files preserve setup commands, conventions, and architectural knowledge between sessions. Guidance on managing Claude Code workflows is directly relevant to this layer.

Test moves quality work into the implementation loop. In addition to unit-test generation, teams create custom eval suites from 20 to 50 real-world tasks drawn from recent development work. This tests whether an agent can complete representative engineering work, rather than measuring coverage alone. Teams building these suites can apply established practices for evaluating AI agents.

Deploy uses multiple AI review layers to check the diff against plan.md and security standards. Git hooks enforce programmatic policies, including whether the plan matches the pull request. Low-blast-radius changes with passing tests can move toward automatic acceptance, while human engineers retain approval authority over regulated, high-risk, and core architectural logic.

Maintain connects production operations back to product planning. Monitoring agents watch logs and metrics, while Claude Tag handles first-response work for CI/CD failures and bug triage. When an anomaly exceeds a defined control bound, the maintenance workflow drafts a new intent.md, restarting the loop instead of leaving incidents in a separate ticket queue.

Rollout guidance for engineering teams

Anthropic recommends keeping CLAUDE.md to roughly one page. Detailed instructions belong in directory-specific rules or Skills, which limits context bloat and makes policy reuse more precise. Teams scaling parallel Claude Code work should begin with two to three concurrent sessions per engineer before adding more.

The governance model also changes. Engineers spend less time performing mechanical edits and routine approvals, and more time evaluating intent, signing off designs, reviewing exceptions, and approving high-risk changes. This aligns with broader work on agent skills versus editor rules and parallel subagents in Claude Code.

For an enterprise rollout, start by versioning intent, design, and implementation plans for one bounded service. Add evals from 20 to 50 real tasks, enforce plan-to-diff checks with Git hooks, and keep human approval mandatory until the repository has evidence that lower-risk changes are safe to automate.