Use cases / Tech leads and senior engineers
Reviewing pull requests written by agents
Agents open pull requests faster than anyone can read them. Read the 10% that decides whether it is safe.
The problem
An agent can produce a 900-line pull request in the time it takes to read 50 lines of it. The usual answers, rubber-stamping or blocking, both fail: one ships the TLS bypass, the other throws away the speed.
What visualdiff does
- Provenance names the agent and whether a person committed after it.
- The review plan puts the security-sensitive hunks first and folds moved or reformatted code away.
- Intent drift flags every area the agent touched that its description never mentioned.
- Policy can hold agent-written changes to a lower size limit, or require a human commit before merge.
In the sample, the plan reads 65 of 582 changed lines closely, skims 190 and skips 327.
Related
- Large refactors that hide a behaviour changeA refactor is supposed to change nothing. Find the one line that does.
- Stopping hallucinated dependenciesSlopsquatting: attackers register the package names models invent. Catch them at review.
- Seeing which running systems a change reachesThe diff says rds.tf. The map says orders-db, the API in front of it, and the queue behind it.
- Knowing whether review is keeping upThe sanity index: is the codebase getting harder to review as more of it is written by machines?