Engineer: LLMs are eroding my software engineering career

A software engineer with ten years of experience has published an essay about how large language models are dismantling the pillars of his expertise. The post appeared on the bearblog platform under the name "human-in-the-loop"; the author writes anonymously.

The first pillar: domain knowledge

Last year he joined a finance company that had embraced AI wholeheartedly, with ChatGPT and Claude Enterprise accounts from day one and an instruction to review and own every line that reached production. His first project was reworking a legacy online payment system. His manager's feedback on the design documents he wrote was blunt: the code pace is fine, but the documents take too long — use more AI. He agreed, and watched the knowledge he had accumulated — trade-offs between implementations, how acquiring works, how to structure idempotency to prevent double charges — become promptable.

The second pillar: debugging and distributed systems

Then they got good at writing code, starting in the second half of 2025 with the Claude Code wave and Codex. Claude 4.5 was not that good: given a stack trace and context — often just a Sentry link with the Sentry MCP enabled — it solved roughly 60% of bugs and sometimes offered plausible but wrong fixes.

The next generations changed the picture. With 4.6, 4.7, GPT 5.5, Opus 4.8 and the Datadog MCP, bugs that previously took one or two full days — race conditions, third-party integration issues, undocumented API edge cases, failures in distributed systems with poor observability — are now solved on the first attempt in about 90% of cases. He is still employed, he notes, because someone has to review the code and steer the agent, but he now calls himself "just another off-the-shelf engineer".

The last pillar: architecture and "taste"

One area still stands: code quality and software architecture — what the industry increasingly reduces to the word "taste". Agents keep codebases poorly organised, producing circular dependencies, duplicate code and SOLID violations. Yet the industry, he argues, is moving toward accepting C- or D-grade codebases, because source code is written for machines to read rather than for humans.

Hiring is changing too: listings that once read "Software Engineer – Area" are now simply "Software Engineer", and a layoff round at his company eight months earlier left several brilliant former colleagues still job-hunting. For the long term he sees no easy exit: returning to university for mathematics and machine learning and then applying to a frontier lab is complicated, since there are no frontier labs in his country and family matters make moving abroad difficult. "Maybe I should consider transforming my woodworking hobby into a profession," he writes.