An anonymous software engineer who says they have spent a decade in backend and fintech engineering has published an essay describing how AI coding agents now resolve the vast majority of complex bugs they previously had to debug by hand. The account, posted on the blog Human in the Loop, is a personal reflection rather than a study, but it offers a detailed first-person look at how one specialist's technical edge has narrowed.
The author, who specialized in areas like PCI compliance, double-entry ledgers, and payment idempotency, writes that domain-specific expertise was the first skill to erode as AI models got better at technical research and design work. They say that after integrating tools such as Claude 4.5 with observability platforms including Sentry and DataDog via Model Context Protocol (MCP) connectors, the agents began resolving what the author estimates as roughly 90% of the complex distributed-system bugs they previously had to debug manually — up from about 60% with earlier tool setups.
According to the essay, multi-service bugs that once took up to two days of manual investigation are now often fixed in a single prompt, aided by root-cause analysis pulled directly from observability data. The author frames this as commoditizing hard-won operational intuition: generic senior developers steering AI tools can now match specialists who spent years learning a particular system's quirks.
The essay also argues that code architecture, the last area where the author felt human judgment still mattered, is being devalued. AI agents left unsupervised tend to produce duplicated code, circular dependencies, and SOLID-principle violations, the author writes, but companies are increasingly willing to tolerate messier codebases as long as tooling can still parse and maintain them — reducing the incentive to spend sprint time on refactoring. The author does caution that unsupervised agents still degrade codebases into "spaghetti code" over time, meaning human review and architectural oversight remain necessary.
Because this is a single practitioner's anecdotal account rather than benchmarked research, the specific percentages should be read as one engineer's estimate of their own experience rather than an industry-wide measurement. Still, the essay reflects a broader anxiety among engineers about AI tools narrowing the gap between specialists and generalists, and about employers' shifting tolerance for code quality once machines are the primary maintainers.