Graphite Agent (formerly Diamond)

You will like Graphite Agent if you already stack PRs and merge through Graphite, because the reviewer turns on in under five minutes with no config. The daily trait is that fixes are one-click suggestions on the stack, so accepting a comment does not mean switching to an editor.

Otherwise you are buying a stacking workflow to get a bot, and the workflow is the bigger change. Pick it as part of the bundle if your team is already sold on stacking. Pick cubic if review is the only problem you have and you want to keep your merge flow as it is.

The risk is write access. The description says the agent reviews, fixes and updates PRs conversationally, and the listing shows multi-file edits with no Docker sandbox. A reviewer that edits is an author with no isolated workspace, and an author that writes into someone else's branch on request is a merge conflict waiting for a human.

The consequence: keep the fix feature as a suggestion, never an autopush, and let the engineer apply it. A documented isolated workspace for the fix step would change this verdict. What it does right: per-file exclusions and a rules tab that reports whether each rule fires, so noise is measurable and removable.

Each request sends 1. pull request metadata, 2. the code changes, 3. related or similar pull requests, 4. related sections of the codebase. That is retrieval, not indexing, and nothing is executed, so the reviewer reasons about code it has never run. The marketing term is high-signal; no benchmark or false-positive rate is published to define it.

The consequence is that the quality claim rests on the retrieval step, which is described only by category, and a reader cannot tell how related is decided. A published precision figure on a fixed PR set would move the score. The observation: the data-processing page is the best architecture doc the product has.

Graphite sells a workflow with switching cost: stacked PRs and a merge queue that a team's habits grow around. Bolting a reviewer onto that seat is defensive, and unlimited AI reviews at the Team tier says the company treats inference as a cost of retention rather than a product to price.

That is the correct call for a workflow company and a bad sign for anyone hoping the reviewer becomes a business on its own. Likely acquirer: GitHub, which has been absorbing this workflow feature by feature and could finish the job. Position: long the company, indifferent to the reviewer.

The demo is a comment on a stacked pull request. Procurement: Team is $40 a user on annual, $50 monthly, so sixty seats is $2,400 or $3,000 a month. SSO, SAML, the SIEM audit log and GitHub Enterprise Server are all Enterprise, which means the tier we can sign is one above the one on the page.

AI features are opt-in and disabled by default, which security will like. GitHub only, so the GitLab team is out. Onboarding is the harder part, because stacking is a workflow change, not an install. Approved with conditions: Enterprise quote in hand and a pilot team that already stacks.

Anthropic and OpenAI are chosen for me, no key of my own, no local model, no MCP client or server, no source. Custom rules and file exclusions are the only dials and they live in a web tab, so my review policy is not in my repo and not in git.

Disabling AI org-wide means emailing support@graphite.com, which is a kill switch with a ticket queue. I use tools I can script; this one I can only click. Nothing to fork, nothing to run at home, and no CLI for the reviewer. It may be a fine reviewer. It is not mine.