JPMorgan Engineer Declares AI Surpasses 15-Year Coding Expertise

The September 2026 blog post from pitermarx.com describes an internal shift at JPMorgan where AI agents handle infrastructure tasks like Terraform that previously required domain-specific expertise. The author notes that prompt-based goal statements now replace iterative coding, with recent model releases removing the need for AGENTS.md constraint files previously required for reliable output. Performance claims lack controlled measurement; no A/B tests compare AI-assisted versus baseline velocity or defect rates at the firm.

Related evidence appears in SWE-bench Verified results from October 2025 showing frontier models resolving 48% of real GitHub issues without human scaffolding, up from 33% in 2024. Anthropic's internal Claude Code metrics released in March 2026 similarly report agent success rates above 60% on multi-file refactors. The post understates these benchmarks by focusing on personal grief rather than tracking commit-level acceptance rates or review time reductions.

Operationally this compresses the value of accumulated syntax and pattern knowledge while raising the premium on specification and verification skills. Inference costs for small models such as GPT-5.6 Luna are projected to fall below $0.01 per million tokens by late 2027, shifting spend from tooling to outcome validation. Large institutions retain guardrails that slow full adoption, yet the post's unmeasured ROI concern remains unaddressed by existing telemetry.

Next phase centers on whether verification layers can scale faster than generation capability. Banks tracking AI-generated code volume against production incidents will determine if the current experimentation window closes or expands.