Coding Agents
Review Is the Job Now
The economics of writing code changed. The economics of being responsible for it did not. Five red flags in an agent-written diff.
The economics of writing code changed. The economics of being responsible for code did not. If an agent writes 80% of the diff, you still own 100% of the outage, and the review is now the only place where that ownership is exercised.
Review the diff, not the summary
Agent-written pull requests arrive with a confident description. The description is generated from the same intent that produced the code, so it inherits the same blind spot. Read the diff. The summary tells you what the agent meant; the diff tells you what will run.
The five red flags
- Unrelated changes — a formatting pass, a renamed variable, a moved file. Scope creep hides defects.
- Deleted or weakened assertions — the fastest way to a green build is to stop checking. This is the single most dangerous edit an agent can make.
- New dependencies — a package added for four lines of code is a supply-chain decision made without a human.
- Swallowed errors — a bare `catch {}` around the failing line. The test passes; the bug is now silent.
- Happy-path only — the feature works when the input is well-formed and the network is up. Nobody's users are that polite.
A green build is evidence that the checks you have passed. It is not evidence that the change is correct.
Make the machine do the boring half
Types, lint, unit tests and a build are not review. They are the floor. Automate them in CI so the human review can spend its attention on the things a machine cannot judge: is this the right change, is this the right place, does this match how we work.
name: verify
on: [pull_request]
jobs:
check:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: actions/setup-node@v4
with: { node-version: 22, cache: npm }
- run: npm ci
- run: npm run lint --if-present
- run: npm test --if-present
- run: npm run buildSmall diffs are a review strategy
Review quality collapses as diff size grows. A 2,000-line pull request gets a skim and a thumbs-up; a 120-line one gets read. If you delegate, delegate in slices: one function, one test, one file. The agent is fast enough that four small runs cost less than one big one, and each one is reviewable.