跪拜 Guibai
← All articles
Frontend · Backend · JavaScript

Anthropic's Playbook for Rebuilding the SDLC Around AI Agents

By threerocks ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Most engineering organizations have adopted AI coding tools but left their surrounding processes—PR reviews, QA gates, change advisory boards—untouched. Those processes now throttle the speed AI unlocks. This playbook provides a complete, auditable blueprint for re-engineering the entire lifecycle so governance keeps pace with agent throughput, which is the difference between a demo and a production system that regulators will accept.

Summary

The traditional SDLC was built for an era when writing code was the slowest, most expensive step. That bottleneck has vanished. Anthropic's Applied AI team now publishes the concrete practices they use internally and with customers to redesign all six phases—Plan, Design, Develop, Test, Deploy, and Maintain—around Claude's agentic capabilities. Each phase produces a version-controlled artifact (intent.md, spec.md, plan.md, code diffs, review results) that automatically triggers the next phase, forming a continuous audit trail.

The playbook is granular: it specifies prerequisites, step-by-step execution, governance controls, and measurable leading and lagging indicators for each practice. Skills encode organizational policies as machine-readable constraints; hooks enforce deterministic guardrails and approval gates; evals run continuously in CI to regression-test agent configuration changes. The end state is a self-propelling loop where production metric excursions spawn new intent.md files that re-enter the pipeline.

Crucially, the framework preserves separation of duties. Agents can execute everything up to a production approval node but cannot cross it alone. Branch protection, managed settings that engineers cannot override, and sandboxed execution with scoped credentials keep the system auditable for regulated enterprises. The final judgment always stays with a human.

Takeaways
Code generation is no longer the bottleneck; the phases before and after development (planning, review, deployment) now throttle delivery because they still run at human speed.
Every phase in the AI-native SDLC ends by committing a version-controlled artifact—intent.md, spec.md, plan.md, code diffs, review results—that automatically triggers the next phase and forms a complete audit trail.
Plan mode in Claude Code forces design review before any code is written; changing direction at that point costs only an edited document.
Skills encode organizational policies (security, brand, compliance) as machine-readable constraints that Claude applies during code generation, while hooks provide deterministic enforcement that cannot be bypassed.
Continuous evals in CI regression-test agent configuration changes—CLAUDE.md, skills, hooks—the same way code is tested, with production incidents permanently added as eval cases.
PR reviews become multi-pass specialized scans (bugs, security, spec compliance) sorted by severity; human reviewers focus only on whether the change matches intent and whether risk is acceptable.
Managed settings for regulated enterprises can lock down permissions, sandbox execution, block credential access, and force all skills and hooks to come from an approved marketplace, with engineers unable to override any control.
The maintenance phase closes the loop: deterministic scripts detect metric excursions, Claude diagnoses and writes findings as intent.md, and the entire pipeline runs unattended until a human approval gate.
Parallel sessions in isolated Git worktrees let one engineer coordinate multiple Claude instances simultaneously, with sub-agents handling recurring verification work in fresh context windows.
Rollback must be the most-rehearsed path in the pipeline, regularly drilled in staging, because the autonomous maintenance loop will invoke it when metrics breach control limits.
Conclusions

The playbook treats the SDLC not as a linear assembly line but as a continuous loop where production incidents automatically spawn new intent artifacts—blurring the line between maintenance and new development entirely.

Anthropic is effectively arguing that the unit of work in software engineering shifts from 'task completed by a human' to 'artifact committed and reviewed,' with the engineer becoming a coordinator of multiple concurrent agent sessions.

The distinction between advisory controls (skills) and deterministic controls (hooks) is a useful taxonomy for any team building agent governance: skills reduce violation probability, hooks make violations structurally impossible.

Requiring that every production incident become a permanent eval case creates a ratchet effect where the test suite grows strictly more comprehensive over time, which is a stronger guarantee than most human-run QA processes offer.

The managed settings example for regulated enterprises reveals how much configuration surface area exists—sandboxing, credential denial, marketplace restrictions, minimum version enforcement—and how few organizations have likely addressed even half of it.

Bidirectional PR review (Claude reviews human code, humans review Claude's review of Claude's code) creates a layered defense that preserves separation of duties while keeping review latency in minutes rather than days.

Concepts & terms
AI-Native SDLC
A software development lifecycle where AI agents are embedded in every phase (Plan, Design, Develop, Test, Deploy, Maintain), each phase produces a version-controlled artifact that triggers the next, and human judgment is reserved for approval gates rather than execution.
intent.md
A Markdown file written by an idea initiator (with Claude's help) that describes a problem, desired outcome, affected systems, and constraints. It replaces traditional PRDs and requirements documents as the entry point into the AI-native SDLC.
plan mode
A Claude Code permission mode that allows the agent to read the codebase but not edit files. Engineers use it to generate and refine an implementation plan (plan.md) before any code is written, making design review happen before code generation.
Skills (Claude Code)
Directories containing a SKILL.md file with frontmatter that specifies when the skill triggers and body text that specifies what policy to apply. Skills are advisory controls that encode organizational standards (security, branding, API conventions) as machine-readable instructions.
Hooks (Claude Code)
Scripts that run before or after Claude executes an action (editing a file, running a shell command). They provide deterministic enforcement—blocking unsafe operations, running formatters, or requiring human approval—and cannot be bypassed by the agent.
Continuous Evals
An evaluation suite that runs in CI to regression-test agent configuration changes (CLAUDE.md, skills, hooks). Each eval is a prompt paired with an acceptance check; production incidents are permanently added as eval cases to prevent recurrence.
Git Worktrees
Independent working copies of a Git repository on separate branches, used to isolate parallel Claude Code sessions so multiple agents can work simultaneously without conflicting on the same files.
Managed Settings
Claude Code configuration pushed by platform or IT administrators that individual engineers cannot edit or override. Used in regulated enterprises to enforce sandboxing, credential denial, marketplace restrictions, and minimum version requirements.
Claude Tag
A Slack integration (public beta) that lets Claude join channels under its own identity to serve as an immediate first responder for incidents, accessing systems via MCP and writing findings back into the SDLC pipeline.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗