Human Review of Agent Diffs
TrialFactorial rates Human review of agent diffs as Trial (2026). Some of us review agent work with revdiff, a terminal UI that opens the diff next to the agent session, collects line-level annotations and returns them to the agent.
When an agent writes most of a change, the developer's main job moves from typing to reviewing. Reviewing in the chat window is slow and imprecise. Reviewing the diff directly works better: the developer reads the changes as a diff, leaves comments on specific lines, and the agent works through the comments as a list.
The pattern often adds an automated pass before the human one: a review by a second model, or by the same agent with a review prompt, that catches obvious bugs first so the person can focus on design and intent.
Updates
Some of us review agent work with revdiff, a terminal UI that opens the diff next to the agent session, collects line-level annotations and returns them to the agent. Before that, a review command or a second agent such as Codex checks the change for bugs. Our lesson: the bottleneck is no longer writing the code but reading it, so we keep agent changes small and focused and review them before they pile up, especially when several agents run in parallel.