Anthropic engineer Addy Osmani has warned that Claude Code can generate software faster than teams can review it. He says the resulting knowledge gaps could make failures harder to fix.
Osmani, a former Google engineer who works on Claude Code, describes the problem as “comprehension debt”. His discussion of the issue circulated among software developers in March.
Traditionally, reviewing code helped engineers understand why systems worked the way they did. It also showed them why earlier developers had introduced safeguards or unusual technical solutions. AI coding tools have changed that process.
Teams can now supervise AI agents that produce large amounts of software in less time. However, Osmani argues that generating more code does not necessarily improve engineers’ understanding.
Automated testing and AI-assisted reviews can help manage the workload. Yet they may leave developers without the knowledge needed to identify hidden dependencies.
As a result, Osmani warns, apparently minor changes could disrupt important parts of a system. Engineers unfamiliar with earlier design decisions may struggle to repair the damage.
Read: Claude Code Backdoor Warning Issued by China Regulator
His concerns have attracted attention during a period of high-profile technology outages. Ageing infrastructure, delayed maintenance and overstretched information technology teams are other concerns discussed in connection with recent service disruptions.
However, the warning does not establish that Claude Code or other AI-generated software caused any particular outage. Osmani has raised particular concerns about software used in healthcare, financial infrastructure and government services.
Read: OpenAI Ethical Hack Used Claude to Reach Private Code Repo
He warned that blaming unreviewed AI-generated code for a serious failure “will not hold up in a post-incident report” when lives or valuable assets are at risk.
With additional input from Yahoo Finance.