By September, more than 40% of pull requests on GitHub started with an agent's first draft, up from 17% in March, and 90% of agent-involved pull requests merged within two weeks, against 83% in February. Among Jellyfish's lead adopters of AI-only review, 9% of pull requests merge with no human review at all. Meanwhile, in Stack Overflow's 2026 survey of more than 30,000 developers, nearly half said they trust AI output when they can verify it easily, and 6.6% would trust it for important decisions.
Read those together and the problem is plain. Verification is the condition of trust, and the approve button no longer proves it happened.
I'm building Deslopping, a GitHub App in private beta, to make that proof cheap. So this is the case for my own product; read it as such. The walkthrough covers the mechanics end to end. This post is the argument, organised around the five objections I hear most often, usually in this order.
"It will slow every pull request down"
It won't, because most pull requests never meet it. Every pull request gets a risk score from five signals: the paths it touches, its reach, the type of change, the file's history and its size. Below the threshold your team sets, Deslopping does nothing at all. Your team also decides which paths count as sensitive, so payments/** can matter more than a README.
Two details matter more than they look. The scoring runs on servers in Germany and does not call a model, so it is fast and predictable. And the AI tool is recorded but never affects the score: a risky change is risky whether Claude Code, Cursor, Copilot or a human typed it.
"Engineers will hate being quizzed"
This is the objection I take most seriously, because a tool engineers resent is a tool they route around. So the questions are few and narrow: two or three, about the specific lines that changed, never about the codebase in general. For someone who read the diff, it is a couple of minutes.
The framing does the rest. A missed question is treated as evidence that the AI wrote code people struggle to follow, not that the engineer failed. When the grader is unsure, the answer passes. After three misses, Deslopping suggests a colleague for a walkthrough rather than a fourth attempt. The point is a second pair of eyes, not a score.
"This is surveillance with a nicer font"
It would be, if it ranked people. It doesn't. There are no per-person scores, no rankings and no leaderboards. Leads see aggregates only, and nothing is shown below five reviewers. Reviewers can read, export or delete their own answers. The one place a name appears in a lead's dashboard is the break-glass log, because an emergency bypass should have an owner.
Code gets the same discipline. Only the changed lines and their surrounding context are sent, never the whole repository; input is capped at 12,000 tokens with secrets stripped first; the questions are written by Anthropic's Claude with training contractually disabled; and question-related code is encrypted and deleted after 90 days by default. Enterprise customers can bring their own Anthropic API key.
"The problem is the AI, not my engineers"
Agreed, and the product is built on that sentence. The only rating in Deslopping is the reviewer's verdict on the AI's work: good, mixed or poor, with a line of explanation if they want. It is recorded against the module and the tool, never against the person.
Over weeks, those ratings become a map of where your AI writes code your team cannot follow, by module and by tool. That is the data most teams lack when they argue about which assistant to standardise on, or which part of the codebase needs better agent rules. A before-and-after view for those rules, and rules delivered as pull requests through a second, opt-in app, come next.
"Production is down and your check is in the way"
Then you break the glass. Authorised people can bypass the gate with a short reason; every use is logged, and the review is still owed within 24 hours. Whether the check fails open or fails closed during a Deslopping outage is your setting, not mine.
Everything else lives where review already lives. The gate is a required status check on the pull request, deslopping/comprehension, next to your build and your linter. When the questions are answered, it turns green and the merge goes ahead. There is no new tool to learn.
The 30,000-foot read
Approval used to mean "I read this." It was never a perfect promise, but it was cheap to keep when humans wrote every line at human speed. Agents changed the arithmetic, not the obligation. Someone still has to understand what ships, and the only parts of this that sit within a team's control are where that understanding is checked and how little it costs. Deslopping does not read your code for you. It makes sure a person did, on the changes where it matters, and turns the friction into a record of how your AI is performing.
It is in private beta with a deliberately small group of design partners. Team pricing is €12 per active reviewer per month, or €10 billed annually, with a three-reviewer minimum; Enterprise starts at €18 for 50 or more reviewers; accounts and dashboard reading are free, and design partners get half off Team for the first year. If your team already ships with Claude Code, Cursor or Copilot and the approve button has started to feel ceremonial, talk to me. The 69-second tour is on the same page.
Sources: GitHub, Octoverse 2026 (agent-started pull requests, 17% to more than 40%; merge rates of agent-involved pull requests, 83% to 90%); Jellyfish, AI engineering trends (9% of pull requests merged with no human review among lead adopters of AI-only review); IT Brief on the Stack Overflow 2026 Developer Survey (October 6, 2026; more than 30,000 responses; trust when verifiable, 6.6% for important decisions); deslopping.app (features, data handling, break-glass, pricing, beta status); How Deslopping works (grader and retry behaviour). DISCLOSURE: I build Deslopping, and this post argues for it. Product details come from deslopping.app and my product specification; the screens are design renders with sample data, not customer data. The product is in private beta and this post claims no customer results. The Stack Overflow figures come via IT Brief's report rather than Stack Overflow's own results page.