AI agents force a rethink of engineering work

Software engineering practices have shifted rapidly since more capable AI agents emerged around November last year, with experienced developers describing a move from hand-writing code toward orchestrating agents, reviewing output, and judging product fit. Linear and Cursor data showed more pull requests, larger changes, and a rise in accepted AI-generated code with less human review.

Meta’s recent Instagram account-takeover outage offers a warning case. Meta engineers attributed the incident to AI-generated, AI-reviewed code, layoffs, and reassignments away from Integrity work toward AI labeling and training tasks. Reported cuts included Instagram’s design team losing 44% of headcount, the Developer Documentation and Support team seeing a 95% headcount reduction, and Instagram’s Trust and Safety Team losing around 50% of its staff to data labeling and layoffs.

Large tech firms are reorganizing workflows around agents. Anthropic is described as heavily reliant on Claude, with ~70-90% of internal code generated by Claude and Claude Cowork built in just 10 days. OpenAI’s Codex team uses AI review, automated fixes, and parallel agents, while Uber has built in-house infrastructure for agent workflows, risk profiles, code review, and migrations. Startups and traditional companies are also investing, including 18,000 Cisco developers using Codex in February.