AI coding tools are still junior

The new generation of AI coding agents can look uncannily capable. They read a repository, edit files, run tests, open pull requests, and explain what they changed in brisk, confident prose. Anthropic says Claude Code can search a codebase, make changes across files, run the test suite, and even commit fixes automatically. That pitch is real enough to be useful. It is also exactly why companies are learning to treat these tools like junior engineers, not senior ones (anthropic.com). That distinction matters because the gap is not in typing speed. It is in judgment. InfoWorld reported on April 7, 2026 that enterprise developers are questioning Claude Code’s reliability on complex engineering work, especially debugging that crosses multiple files and system boundaries. A separate TechRadar report, published April 6, argued that AI agents should be trusted only as “junior engineers” and kept inside tight governance, with limited access and mandatory review before anything reaches production (infoworld.com; techradar.com). Benchmarks help explain why the sales pitch and the field reports can both be true. SWE-bench Verified, a widely watched test of real GitHub issues, was built as a human-validated set of 500 software tasks. It measures whether a system can produce a patch that actually resolves a real bug or feature request in a live codebase. The leaderboard shows rapid progress, but the benchmark itself exists because software work is messy enough that toy coding tests are not good proxies for reality (swebench.com; github.com). And even those benchmark gains do not cleanly transfer to day-to-day engineering. In a July 10, 2025 randomized trial, METR studied 16 experienced open-source developers working on their own mature repositories, with real issues drawn from projects averaging more than 1 million lines of code. When those developers were allowed to use AI tools, mostly Cursor Pro with Claude 3.5 or 3.7 Sonnet, they took 19% longer to finish the work. The tasks averaged about two hours. The surprising part is that the developers themselves thought AI had sped them up. It had not (metr.org). That result fits a broader pattern. Stack Overflow’s 2024 developer survey found that 76% of respondents were already using or planning to use AI tools in development, yet trust remained shaky. Only 43% felt good about AI accuracy, and the company’s own summary said developers saw improvements in quality of time spent, but not necessarily in time saved. Adoption is no longer the question. Reliability is (stackoverflow.blog; survey.stackoverflow.co). The quality problem is not just about wasted hours. It is also about what gets shipped. Veracode’s 2025 GenAI Code Security Report tested more than 100 models across Java, JavaScript, Python, and C#. In 45% of tests, the generated code introduced risky security flaws. Bigger and newer models did not solve that problem. Google’s 2025 DORA report reached a different but related conclusion: AI acts as an amplifier. In strong engineering organizations, it can magnify good practices. In weak ones, it magnifies the mess (veracode.com; research.google). So the practical lesson is not to stop using the tools. It is to use them where junior help is actually helpful. Let the agent scaffold a feature, write boilerplate, generate tests, search a large codebase, or draft a migration plan. Keep humans on the work that depends on architecture, invariants, failure modes, and production judgment. Anthropic’s own marketing now describes engineers as the people who focus on architecture, product thinking, and orchestration while the agent handles execution. The machine writes fast. The human still decides what must not break (anthropic.com).