LLM Agents Are Breaking GitHub PRs Through Accumulated Architectural Debt.

The volume of code that large language model agents are now producing is staggering, and it has exposed a mismatch between how that code flows into pull requests and how humans actually review it. When an agent fills a substantial share of a PR, the tests may still pass, but the architectural integrity often degrades into tangled cross-module dependencies and a breakdown in separation of concerns, because the optimization target is the immediate ticket rather than the long-term health of the codebase.

AI-powered review tools such as CodeRabbit, Copilot, and Claude Code scanning a PR are competent at flagging missing null checks or style violations, yet they struggle to detect duplicate logic spread across separate directories or a design pattern that conflicts with the rest of the system. The interface layer has become the most significant friction point: GitHub's PR UI was already awkward for human-authored changes of modest size, and it becomes nearly unusable when an agent commits a large block of boilerplate and logic in one go, producing a noisy diff that mixes human comments with automated bot reviews and invites the practice of pasting AI suggestions into comments without validation, which in turn leads reviewers to rubber-stamp oversized diffs rather than parse them.

Sustainable AI-assisted workflows therefore benefit from separating the functional check from the architectural check, and this suggests the PR process itself may need to be reconsidered. For high-velocity teams, the practical questions revolve around whether smaller, more atomic PRs are being adopted to manage volume, and whether AI-generated noise can be filtered out before it reaches a human reviewer; a real-world deployment strategy for human-in-the-loop reviews that avoids total burnout is becoming essential rather than optional.

Web agents also reveal a structural assumption baked into HTTP, HTML, CSS, and JavaScript: every layer assumes someone is looking at a screen, which breaks down when the consumer is a programmatic agent rather than a person reading a page. The thesis gaining traction is that the web needs an execution layer for agents, not just a reading layer, and projects like Rover are beginning to build it; understanding how agents consume content is essential to designing protocols that work for all of them. Text-Based Agents such as Claude, ChatGPT, and Gemini chat access the web through tool calls to search and fetch APIs, while some agents like Claude Code and OpenCode send Accept: text/markdown to request markdown when available. Training crawlers like ClaudeBot and GPTBot identify themselves via User-Agent strings, but inference-time fetches often use generic or internal User-Agent strings with no crawler identity, making it harder to distinguish automated traffic. CUA Agents, or Computer Use Agents such as Claude CUA and OpenAI Operator, operate through screenshots and therefore interact with the rendered interface directly, requiring different handling than text-based fetches.