Anthropic published an AI-native SDLC playbook. The interesting part isn't the six stages, it's what replaces line-by-line review
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
Claude 相关模型动态已经出现，适合跟踪能力变化、生态影响和后续可用性。
Anthropic 发布了一份 AI 原生 SDLC 剧本，指出当 AI 代理生成大部分代码时，传统的逐行代码审查将不再适用。Faros AI 的数据显示，在涵盖 10,000 名开发者和 1,255 个团队的范围内，高度采用 AI 的团队合并的 PR 数量增加了 98%，但审查时间增加了 91%，平均 PR 大小增加了 154%。…
Link: https://claude.com/blog/the-ai-native-sdlc-playbook
Short version: six stages (plan, design, build, test, deploy, maintain), each one ending in a committed markdown artifact. intent.md, spec.md, plan.md, then the PR and the incident record. Agents generate and verify, humans approve at gates, and a hook can block an action until a named person signs off.
The part I keep coming back to is their claim that reviewing every line by hand stops working once agents write most of the diff. Faros AI's telemetry (10,000 devs, 1,255 teams) lines up with that: high-AI-adoption teams merge 98% more PRs, review time up 91%, average PR size up 154%. DORA 2025 gets there from another angle, throughput up and stability down.
So the bottleneck moved from writing code to deciding what to write and checking what came back. Same shape as 1957, when people argued a compiler could never match hand-written assembly. It didn't have to match. It was abundant.
Open source is already converging on the same loop: GitHub's Spec Kit (specify, plan, tasks, MIT), AGENTS.md sitting in 60k+ repos, and someone has already packaged the playbook itself as a Claude Code and Codex skill.
Question for people running agents at work: what actually replaced line-by-line review for you? Evals in CI, a verifier subagent with a fresh context, hooks on protected paths, something else? And what still slips through?