A senior engineer watches LLMs erase the skills that defined his career

A backend engineer with a decade of experience, much of it specializing in payments and finance, describes watching three pillars of his expertise collapse in succession as LLM tools matured. First went domain knowledge: models trained on the public corpus of articles about PCI compliance, idempotency, and ledger design could draft architectural design docs and reason about payment flows nearly as well as he could. Then went debugging — the skill he was sure would remain human. With agentic coding tools, Sentry and Datadog MCP integrations, and frontier models from Claude 4.5 through Opus 4.8, he reports that roughly 90% of distributed-systems bugs that once took him a day or two to track down are now one-shotted by the agent.

The last pillar he claims is still standing — code quality and architecture — is being devalued rather than automated. Agents produce sloppy, duplicative, dependency-tangled code, but the industry seems to have decided that a C-grade codebase is acceptable because the code is increasingly written for machines to read, not humans. He frames this as an economic problem more than a technical one: when every senior engineer becomes an interchangeable generalist steering an agent, the price of that labor falls just as demand softens.

The piece is notable less for any specific technical claim and more as a candid first-person account of an experienced engineer recalibrating his sense of professional worth in real time. He stops short of prescribing a response, ending on uncertainty rather than advice.