Anthropic delivered an uncomfortable judgment this week: coding is no longer the bottleneck of software R&D. When AI compresses development implementation to a few hours, what actually holds traditional companies back is the approval, review, and compliance processes designed for a much slower pace.
What deserves attention isn't "how well AI writes code," but a practical playbook: how to systematically transform the six R&D stages (planning, design, development implementation, testing, deployment, maintenance). After reading, we believe this is a reminder for enterprise managers—not just a toolbook for programmers.
What this is
Traditional software R&D is divided into six stages, with each stage relying on documents, tickets, and approval records to hand off information. Its hidden assumption is that "coding is the slowest, most expensive, and most error-prone" part—so various control points are designed around it.
The AI-native model breaks this assumption and does three things:
First, at the end of each stage, the corresponding artifacts (requirements document intent.md, design document spec.md, plan document plan.md) are written into version control (i.e., the code repository). The next stage reads them directly to start—no more human chasing for handoffs.
Second, AI is embedded in every step. Planning is aggregated by Claude from raw materials; testing runs as continuous evaluation throughout the process, no longer a one-time check at stage boundaries; deployment is reviewed by multi-layer agents for specialized audits.
Third, human attention is reallocated. When AI executes, governance requirements are triggered immediately by hooks (i.e., automated policy-validation hooks); humans only need to review the nodes flagged by agents.
How the industry sees it
The playbook has strong recognition in engineering circles. A tech lead at an AI applications company told us: "Version-control-driven workflows, writing intent as machine-readable artifacts into the repository—we've been doing this internally for a while, but no one has systematically documented it."
Opposition exists as well. A chief architect at a fintech company raised two reservations: the playbook assumes companies have "mature product owners" who can write intent documents with judgment—but at most Chinese companies this role is rare, with project managers mostly paraphrasing requirements; the playbook barely discusses AI explainability in regulated scenarios. When regulators ask "why was this code written this way," AI cannot provide audit-grade answers, which in finance and healthcare is harder than what the playbook addresses.
A more hidden risk: the playbook assumes companies accept "explosive growth in code output." A former security lead at a major internet company warned that if security staffing ratios aren't adjusted, the faster AI writes code, the more unreviewed code goes live. The playbook does mention this, but offers no answer on how companies should reset their staffing ratios.
Impact on regular people
For enterprise IT: If your company already uses AI to write code, the first thing worth checking isn't AI capability, but the cadence of approvals, meetings, and cross-team handoffs. These processes were calibrated to "humans writing code" speed—when AI doubles output, this becomes the new bottleneck.
For individual careers: Developers' work focus shifts from "writing" to "defining." Writing intent, judgment, and constraints clearly will be more valuable than typing code itself. Those holding product decision-making power are the beneficiaries; pure executors are an early warning sign to prepare for.
For the consumer market: Short-term product iteration will speed up, but the apps and enterprise internal systems users interact with are most likely not yet AI-ized in their R&D process—the update cadence you feel as a user won't sync with AI's coding speed.