Stop reading diffs.
Approve intent. Trust evidence.
AI made writing code and preparing documents fast. Review did not get faster. It got bigger. Here is how to move the human check to where it still adds value.
Citește în română →The bottleneck moved
Most teams now use AI every day. Code gets written in minutes, invoices get read in seconds, reports get drafted before the meeting starts. And yet, in many companies, work does not reach the customer much faster than it did a year ago.
The reason is simple: the steps around the work stayed the same. A person still reads every line before anything moves on. When the amount of output grows and the reading does not, the queue in front of the reviewer grows instead.
More output plus the same review process equals a longer queue. Adding reviewers helps for a month. Then the agents get faster again.
Why reading every line no longer works
Reading a diff answers the question “does this code look reasonable?”. That was a useful proxy when a person wrote every line slowly and the reader knew the author. With agents, the proxy breaks in three ways:
- Volume. An agent can produce in an hour what a reviewer can read carefully in a day.
- Plausibility. Generated code and text look clean. Mistakes hide in what was not written: a missing edge case, a wrong assumption about the data.
- Wrong question. The question that matters is not “is this line fine?” but “does this do what we intended, and how do we know?”.
So we suggest changing what the human checks, not how hard they check it. Two principles sit behind everything we do.
Humans approve intent.
Stop checking lines of code. Verify that code and tests express the intended behavior.
Confidence comes from evidence.
Design specs, benchmarks and automated tests establish trust, not reading diffs.
What “approving intent” looks like
Before an agent starts, a person writes or approves a short intent note: what should change, for whom, what must not change, and three to five examples of correct behaviour. That note is the contract.
When the work comes back, the reviewer does not start with the diff. They check two things: does the intent note still say what we want, and do the tests encode those examples? If yes, the change moves on. If no, the fix is in the intent or the tests, not in a comment on line 214.
What “evidence” looks like
Each change arrives with an evidence pack that a machine produced and a person can read in two minutes:
- the intent note it implements, with a link;
- which tests were added or changed, mapped to the examples in the note;
- benchmark results against the previous baseline (speed, cost, accuracy);
- static checks, security scans and dependency checks, all green or explained;
- a risk tier that decides how much human attention the change gets.
People still read code. Just not all of it.
This is not “stop reviewing”. It is “review where reading still finds problems that evidence cannot”. We use simple risk tiers:
Once tests and checks are in place, most routine changes can live in tier 1. That is where the waiting time disappears.
The same idea works in the back office
Replace “diff” with “invoice batch”. An agent matches 400 invoices to orders. Nobody should re-check 400 lines. A person approves the matching rules (the intent), and the agent delivers evidence: totals reconcile with the ledger, a list of the 12 exceptions it could not match, and a random sample of 10 for spot checks. The team handles the exceptions. The rest moves on the same day.
How to start this month
- Pick one flow where work waits for review: pull requests, invoice approval, monthly reporting.
- Measure it for two weeks: how long work waits, how often review finds a real problem.
- Write an intent template of half a page and use it for every new change in that flow.
- Define the evidence pack and automate what you can: tests, checks, totals, samples.
- Agree on risk tiers and let tier 1 move on evidence alone. Compare the numbers after a month.
None of this needs a new platform. It needs a decision about what people are responsible for checking.
Where does work wait in your company?
Take the 12-question scorecard to see where to start, or book a free 30-minute call. We will look at one flow with you.
Sources
- DORA, State of AI-assisted Software Development 2025, summarised by InfoQ: infoq.com.
- F. Chen and J. Stratton, working paper on AI coding tools and code review, as reported by TechRadar (October 2026): techradar.com.
How this article was made: drafted by KYA’s agentic setup from the sources above, then reviewed and approved by a person before publishing.
