What would need to happen for AI to replace junior software engineers

A recent analysis explores the conditions under which agentic coding could replace junior engineers, concluding that three of four key requirements are not met. The fourth, however, is a cause for concern as it does not rely on the others. The study references METR’s time-horizon work, which tracks the length of tasks that models can complete with 50% accuracy. From 2019 to 2025, this horizon roughly doubled every seven months, and recent data suggests the rate of growth has accelerated. However, the methodology of these benchmarks has been called into question, as they focus on a single aspect of the job that has been stripped of its complexity. OpenAI recently stopped reporting SWE-bench Verified results, citing issues with the dataset. They found that 59.4% of audited problems had flawed test cases that rejected functionally correct solutions, and that models could reproduce exact problem details, indicating training contamination. When tested on a harder, less contaminated set, performance dropped significantly. Newer benchmarks like Terminal-Bench and long-horizon evolution benchmarks are being developed to address these issues. A randomized controlled trial by METR involved 16 experienced developers working on 246 real tasks. AI was allowed or disallowed at random, with screen recordings capturing the process. Developers expected a 24% speedup, but actual results showed they were 19% slower. While the tools used were early-2025 models and the sample was small and specific, the study highlights a persistent perception gap in productivity. Survey data from Stack Overflow indicates that 84% of developers use or plan to use AI tools, while 46% distrust their accuracy. Google’s DORA research found that 90% of professionals use AI at work, with over 80% believing it improves productivity, though 30% report little or no trust in AI-generated code. These findings suggest AI acts as an amplifier rather than a replacement. Employment trends show a divergence between younger and older workers in AI-exposed occupations. Stanford’s Digital Economy Lab found a 15% shortfall for 22-25 year olds in July 2025, rising to 19% by June 2026. The adjustment is occurring through reduced hiring rather than layoffs. The study notes that employment has fallen among young workers in codified knowledge roles, while it has increased among experienced workers in tacit knowledge roles. The analysis concludes that agentic coding is not replacing junior engineers but rather the tasks traditionally assigned to them. The bottleneck lies in verification, context, and judgment, areas where the frontier models are still far from matching human capabilities. Hiring decisions are based on outdated benchmark numbers that the original creators have retired. The firms that will likely succeed in the future are those running the 'boring experiment' of hiring juniors with AI tools and measuring their progress toward senior judgment. The author remains cautious and is watching for signs that three key conditions could be met, which would warrant a reevaluation of the current conclusions.