The bugs aren't in the AI. They're in the seams.
An empirical analysis of 3,800+ bugs across Claude Code, Codex, and Gemini CLI (arXiv:2603.20847) finds that 36.9% stem from API, integration, or configuration errors. Not hallucination. Not incorrect code generation. The plumbing between the AI and the tools it invokes.
Note (2026-11-10): I checked the full text of arXiv 2603.20847 (both the first and the current version). The paper studies 3,864 bugs (2,343 for Claude Code, 329 for Gemini CLI, 1,192 for Codex). It reports that 37.3 percent stem from API, integration or configuration errors, and that tool invocation accounts for 37.6 percent and command execution for 25 percent; the essay's 36.9, 37.2 and 24.7 percent do not appear in either version. The 67 percent functionality figure, the 18.3 percent API errors and the 12.7 percent command failures do appear.
The distribution tells the story: 67% of bugs are functionality-related, but the functionality failures cluster at tool invocation (37.2%) and command execution (24.7%). The AI produces the right intent. The intent fails to reach the tool correctly, or the tool's output fails to reach the AI in usable form. The error is at the boundary, not in either system.
Users report API errors (18.3%), terminal problems (14%), and command failures (12.7%) as the most frequent symptoms. These are infrastructure complaints, not intelligence complaints. The user isn't saying “the model suggested wrong code.” They're saying “the model tried to do the right thing and the execution failed.”
This maps onto a general pattern in composite systems: when you compose a capable reasoner with capable tools, the performance bottleneck moves to the composition layer. The AI can reason about code. The terminal can execute code. The bug is in the handoff between reasoning and execution — malformed commands, misread tool outputs, configuration states that neither system fully controls.
The implication for AI coding tool development: improving the AI's reasoning quality has diminishing returns once the integration seam becomes the dominant failure mode. The next increment of reliability comes from making the seams more robust — better error handling at tool boundaries, more informative feedback from execution environments, stronger contracts between the AI and its tools.
The smartest worker can't fix bad plumbing. The seam is the system.