guide
ai code review agents that propose, not nitpick
most ai review bots leave comments on your pull request: twenty speculative concerns, a few nitpicks, and now and then something useful. diff social turns that around. agents work in a room beside the people reviewing, answer the question they were asked, and hand back a proposal that a person accepts or rejects.
why ai code review feels like noise
the complaints are consistent wherever developers talk about review bots: too many comments, too many nitpicks, confident and wrong, and impossible to tell to stop. the cause is the shape, not the model. a bot that reviews every pull request on its own must guess what matters, so it says everything, and every guess becomes a comment someone has to read and resolve.
agents propose; people decide
in diff social, an agent never starts on its own and never writes to your pull request. instead:
- a person selects the lines in question and asks one or more agents about them;
- each agent works in the open, in the same stream as the people in the room, so everyone sees what it was asked;
- it answers with a proposal: an analysis, or the smallest patch that does what was asked, as a unified diff;
- a person accepts or rejects the proposal. nothing is merged, approved or commented on github by an agent.
the result is review where the agents do the tedious reading and people keep the judgement, including the part of review that is really teaching.
the agents in a room
| agent | what it does | uses a model |
|---|---|---|
| flags | scans added lines for secrets, focused or skipped tests, debug output, suppressions and todos | no, deterministic |
| history | reads the repository's history: the file's risk, recent events, and files that usually change with it but are missing | no |
| semantics | for json, yaml, toml, xml, ini and env files: which keys changed, were added or removed, and whether it is only formatting | no |
| reviewer | reads the selected lines and their context and reports risks with line references | yes |
| patcher | proposes the smallest patch that achieves what you asked, as a unified diff | yes |
any other agent can join if it speaks the a2a 1.0 protocol: see a2a for code review.
is ai code review worth it?
when it answers questions people asked, on lines people chose, yes: that is reading done for you. when it comments unprompted on every pull request, the cost of reading its comments often exceeds what it finds. diff social only does the first kind, and every paid answer is charged against a monthly budget that stops at its ceiling.
can ai replace code review?
not in diff social, by design. an agent can propose a patch; only a person can accept it, and people remain responsible for what ships. agents make a review faster; they do not make it optional.
reviewing together, live
everyone who opens the same pull request is in the same room: you see who is reading, messages arrive as they are written, and each one points at the lines it is about. when the branch moves, those anchors follow the code, or say plainly that the lines are gone.
try it on a pull request
paste any public github pull request, commit or compare on the home page, and you are in its room.
open a room