Developers have been promised, repeatedly, that artificial intelligence would make code review painless. Instead, it's become something of a cautionary tale.
GitHub disabled a Copilot feature in early April after users noticed the AI had developed an odd habit: slipping promotional language into pull request summaries. The kind of thing that makes you wonder who—or what—is really writing your documentation. For engineering teams already moving at breakneck speed, the episode crystallized a lingering anxiety. AI can scan code faster than any human. But should we trust what it finds?
Stage, a startup fresh out of stealth, thinks the answer is no. Or rather, not yet. Not without guardrails.
The company emerged from Y Combinator's latest batch this week with a deliberately modest pitch: Instead of automating code review, why not just make human review less excruciating? Its core product breaks pull requests into logical "chapters"—coherent chunks of related changes—rather than forcing reviewers to slog through a chronological mess of commits. Think of it as a table of contents for your code diff.
It's automation, sure. But automation designed to help humans think, not replace them.
The Numbers Behind the Skepticism
That positioning isn't just philosophical hedging. A research study found AI suggestions increased code complexity and size more than human suggestions, examining 278,790 code review conversations across 300 GitHub repositories. Human reviewers, for all their slowness, still bring context that algorithms miss—gut checks about testing, documentation, knowledge transfer. The soft stuff that doesn't show up in a diff.
Charles Pan and Dean Stratakos, Stage's two-person founding team, saw this dynamic play out firsthand. Both Stanford computer science graduates, they spent time at Five Rings, a quantitative trading firm where Stratakos worked on AI initiatives and built a coding agent from scratch. Close proximity to the technology apparently bred a dose of realism. "Teams are merging changes they don't fully understand," the company's Y Combinator profile warns, pointing to AI-accelerated velocity as both symptom and disease.
Speed without comprehension, in other words, isn't really progress.
Chapters, Not Chaos
Here's how Stage works: You open a pull request. The platform analyzes the diff, clusters related changes together, and presents them as structured chapters. Each time you push a new commit, the chapters regenerate automatically.
It's a departure from git's chronological commit history, which—let's be honest—is often a archaeological dig through someone's thought process. "Git commits are temporal and often messy," Stratakos explained in the Hacker News thread announcing the launch. Chapters operate at a higher level of abstraction, reordering changes into logical flow rather than temporal sequence. The team calls it "auto stacked diffs," though that phrase might mean more to engineers than to the rest of us.
Early adopters are apparently treating chapters as the fundamental unit of review. Pan mentioned in the same thread that the roadmap includes marking chapters as viewed and allowing comments on individual sections—turning the whole thing into a structured workflow. The platform also generates pull request summaries and plans to integrate with project management tools like Linear, pulling in the "why" behind code changes, not just the "what."
Manual Override Included
Perhaps more telling than what Stage does is what it doesn't do. The product allows for human intervention at every turn. The founders are considering a CHAPTERS.md file that would let developers manually specify how their pull request should be structured. You can regenerate chapters on demand. They're actively asking users to flag cases where the auto-generated breakdown misses the mark.
It's a bet that the market wants structure, not surrender.
And the market Stage is entering? Crowded doesn't quite capture it. GitHub Copilot has been steadily expanding its code review features since entering public preview early last year, with updates rolling out through the year. GitLab launched Duo Code Review last fall. CodeRabbit offers tiered pricing for automated review. Google's Gemini is adding code review to its Conductor CLI extension. Every few weeks, it seems, another player announces AI-powered code analysis.
When a Hacker News commenter suggested GitHub or GitLab could simply replicate Stage's approach, Stratakos pointed to Graphite—a stacked diffs tool whose features the incumbents didn't copy for years, until Cursor acquired it late last year. Speed of execution, he argued, creates its own moat. Whether that holds true when you're competing against GitHub's distribution and resources is... an open question.
The Contrarian Bet
Stage hasn't disclosed pricing yet, though one unverified claim on Hacker News pegs it at 50% more than Claude Pro. The website requires JavaScript to load, and most details remain scattered across launch threads and a sparse Y Combinator directory listing. For a product this young, that's probably fine. The substance is in the idea.
And the idea—automation that enhances human judgment rather than bypassing it—feels almost contrarian in a moment when every tool is racing to do more of your job for you. The research findings linger: AI can screen at scale, but oversight still matters. Context still matters. Understanding what you're shipping, perhaps more than the founders expected, still matters.
Whether chapters become the new unit of code review or just another abandoned abstraction will come down to execution. But in a market glutted with AI agents promising frictionless development, a tool designed to help humans do their jobs better might be the bet that actually sticks.
At the very least, it won't slip marketing copy into your documentation when you're not looking.
