From Writing Code to Reviewing It: A New Center of Risk
AI code generation is a development model in which large language models write most of the source code while human engineers focus on specifying requirements, reviewing diffs, and validating behavior instead of manually composing every function, which shifts the main engineering risk away from typing errors toward gaps in tests, review discipline, and deployment checks. Anthropic has become a prime example of this shift. According to Anthropic, Claude now writes more than 80% of the code merged into its production systems, and employees are merging around 8x more code per active contributor than 18 months ago. Engineers still decide which work to pursue and what to merge, but their daily work is changing: less time in editors, more time reading AI-authored pull requests, strengthening tests, and managing rollout gates. The question is no longer whether models can generate code, but whether teams can keep review and approval processes reliable at this speed.
Dynamic Workflows and the Rise of AI-Assisted Development Loops
Anthropic’s Dynamic Workflows show how AI-assisted development is moving past single prompts into long-running, multi-agent engineering loops. Claude can now create orchestration scripts, split tasks into subtasks, coordinate many specialized agents in parallel, and validate intermediate results before returning a final change set or analysis. This makes it suitable for complex software engineering tasks such as large refactors, security audits, and infrastructure-wide investigations that once demanded intensive coordination by human leads. A new ultracode setting lets Claude decide when to activate workflow-style execution automatically, saving progress so long tasks can resume after interruptions. Internally, these capabilities feed the same loop that builds Claude itself, with AI agents helping design, implement, and test new features. As Dynamic Workflows mature, they push more work to machines and widen the gap between code creation and the slower, more human-limited phases of review and approval.
Productivity Gains: 8x More Code, 80% Claude Production Code
Anthropic’s internal metrics show how far AI-assisted development has gone in a short time. The company reports that average lines of code merged per active contributor per day have reached 8x the pre-2025 baseline, after years of flat output. The growth closely tracks Claude’s public and internal releases, including Claude 4 and the Mythos Preview models. At the same time, Claude’s share of production code has climbed above 80% of merged changes in some recent months. One notable data point is that, by late 2025, some Anthropic engineers had stopped opening code editors for many tasks, instead asking Claude Code to generate first drafts and then editing the result. For software engineering productivity, this suggests an emerging pattern: humans increasingly steer work and refine AI drafts, while the raw typing and boilerplate fall to automation.
When AI Gets It Wrong: The rsync Backup Failure Warning
The rsync backup incident underlines why review risk matters as much as AI code generation capability. After rsync 3.4.3, some users found that incremental backups no longer behaved as expected, in some cases failing on anything other than full backups. While investigating, users noticed dozens of commits attributed to “tridge and claude”, tying Anthropic’s assistant to a widely used synchronization tool. That discovery triggered a sharply worded GitHub post titled “Please Do Not Vibe Fuck Up This Software” and sparked debates across Reddit and Hacker News. Rsync creator Andrew Tridgell responded in a Medium post, arguing that many critics misunderstood how AI tools were applied. Even so, the episode highlights a key point: failures are now often review failures, not generation failures. If AI-assisted development touches critical utilities, weak testing or rushed review can turn a subtle logic mistake into broken backups for countless systems.
Designing Code Review Automation for an AI-First Future
As Claude and similar tools generate more of the world’s software, code review automation becomes central to risk management. Anthropic emphasizes that engineers remain in the loop, with audit trails, security checks, rollback paths, and human approval required before AI-written code reaches production. Yet an 8x increase in code volume means traditional review practices will struggle unless they are augmented by smarter tools. Likely responses include AI-assisted code review that flags risky patterns, expanded test suites tied into continuous integration, and stricter deployment gates for security-sensitive or infrastructure components. Teams also need clear policies about where AI code generation is acceptable, how to attribute AI contributions, and how to respond when an AI-assisted commit is implicated in an incident, as in the rsync case. The future of software engineering will be defined less by who types the code and more by how we test, review, and monitor what machines produce.
