The industry is suffering from a collective delusion. We’ve treated the manual Pull Request (PR) as a sacred ritual of "quality," yet it has become the single greatest bottleneck in the modern delivery pipeline.
As we move into an era of Generative AI and agentic workflows, the traditional code review isn't just slow—it’s redundant. If you want to maximize delivery efficiency, you need to stop policing lines of code and start orchestrating intelligence.
The "Cargo Cult" of Manual Review
Most teams perform code reviews because of "cargo culting"—they do it because they’ve always done it. When you ask a senior engineer why they are reviewing a specific PR, the answers are usually vague: "To catch bugs" or "To ensure quality."
The reality? Humans are statistically terrible at spotting bugs in a diff. We are great at spotting missing semicolons or naming inconsistencies—things that a linter should have caught before the code was even committed.
The Efficiency Delta: If your delivery is stalled for 24 hours waiting for a human to "LGTM" a change that passed CI, you aren't ensuring quality; you're manufacturing latency.
Gen AI: The Ultimate "Rubber Duck"
The argument for manual review often centers on the need for a "second pair of eyes." In the past, this meant Pair Programming or PRs. Today, Generative AI has fundamentally shifted this landscape.
Gen AI serves as a sophisticated rubber duck. It allows for "Embarrassment-Driven Refactoring"—the process of iterating with an LLM to tighten a domain model or simplify logic before a human ever sees it.
How AI Supplements Pairing:
• Technology Adjacent Execution: AI excels at writing the "glue" code—the Node.js boilerplate, the CSS, the WordPress hooks—that you understand but don't want to spend cognitive cycles on.
• Iterative Feedback Loops: Instead of waiting for a scheduled pairing session, you can "pair" with an agent in real-time, challenging its trade-offs and evolving the design iteratively.
• Contextual Scanning: AI is better than humans at scanning for "janky" patterns across thousands of lines of code, identifying global variable leaks or CSS duplication in seconds.
Three Reasons to Stop Reviewing (And What to Do Instead)
To achieve high-leverage execution, you must categorize code changes by their "In Order To" and automate the path to production.
| Category | The Old Way (Manual) | The New Way (Automated/AI) |
|---|---|---|
| Policy | Manual gatekeeping for "rules." | Automated Enforcement. Linters, static analysis, and AI agents verify compliance. |
| Knowledge | Blocking PRs to "share info." | Asynchronous Show & Tell. Merge first; share the delta in a team feed for awareness. |
| Critique | Nitpicking variable names. | Design Evolution. Use AI to rubber-duck the architecture, then ship. |
The Efficiency Thesis
If we are generating ten times the code using AI agents, we cannot require ten times the human review. That math doesn't scale.
Delivery efficiency is found by inverting the power structure:
- Automate the "What": Use CI/CD and AI to verify that the code works and follows the rules.
- Focus Humans on the "Why": Reserve human intervention for high-level architectural critiques—not for checking if a feature branch is "ready."
- Trust the Pipeline: If the tests pass and the AI-driven "policy" check clears, the code should be in production.
The future of software engineering isn't about reading more code; it's about writing better prompts and building more robust automated gates. The manual PR is dead. Long live the pipeline.
[ START ]
|
v
+-------------------------+
| Write Code with AI | <-----------+
| (The "Rubber Duck") | |
+-------------------------+ |
| |
v |
+-------------------------+ |
| Automated Policy Check | |
| (Security/Lint/AI Scan) | -- FAIL ----+
+-------------------------+ |
| |
PASS |
| |
v |
+-------------------------+ |
| Automated Tests | |
| (Unit/Integration) | -- FAIL ----+
+-------------------------+
|
PASS
|
v
+-------------------------+
| PRODUCTION |
| (Immediate Deploy) |
+-------------------------+
|
v
+-------------------------+
| Async Knowledge Share |
| (Post-Merge Review) |
+-------------------------+
Summary for the High-Agency Engineer
Stop using humans as expensive compilers. If you are still blocking merges for "sanity checks," you are holding the tool wrong. Leverage Gen AI to handle the "adjacent" complexity, automate your policy, and focus your team on the vector of the work, not the syntax.
Top comments (0)