An apply and typecheck gate
DiffJudge applies every candidate diff in a throwaway worktree and runs the project’s typecheck. The verdict comes from git and the compiler, not from a model’s reading of the diff. How the judges work.
Read from each vendor’s own pages on 2026-10-21
Several agents on one task is not new. The gate is the difference.
Cursor, Codex, GitHub, Conductor, OpenHands and Blackbox AI all run more than one agent on a task. Cyberine’s own claim is narrower: it sends one task to the agents you already log into, from different vendors, applies and typechecks each diff, and lands the judged winner or opens it as a PR.
| tool | Several agents on one task | Other vendors’ agents | Judges the candidates | Lands the result | Uses your own subscription login | Where it runs |
|---|---|---|---|---|---|---|
| Cursor /best-of-n | yes, one worktree per run Cursor worktrees docs | none found; models from several vendors inside one Cursor agent Cursor home page | no: “compares runs only” Cursor worktrees docs | partial: you pick, then apply the worktree yourself Cursor worktrees docs | none found; plan pool plus API-rate usage Cursor pricing | local worktrees and cloud agents Cursor cloud agent docs |
| Codex cloud --attempts | yes, 1 to 4 attempts of one task codex cloud-tasks cli.rs | none found; OpenAI models only Codex models docs | no scoring in the CLI; a third-party post says it does not rank attempts third-party post | partial: apply one attempt you choose codex cloud-tasks cli.rs | partial: your ChatGPT plan, no other vendor’s login Codex pricing docs | OpenAI cloud Codex cloud docs |
| GitHub Agent HQ | yes, assign an issue to Copilot, Claude and Codex GitHub blog | yes, Claude and Codex (preview) GitHub plans | no scorer found; you compare the draft pull requests GitHub blog | draft pull requests; no pick-one step documented GitHub third-party agents docs | not documented; runs on the Copilot subscription and AI credits GitHub third-party agents docs | GitHub cloud, GitHub-hosted repositories GitHub cloud agent docs |
| Conductor | yes, parallel workspaces Conductor home page | yes, Claude Code, Codex, Cursor, OpenCode Conductor docs | not documented; each workspace has its own diff Conductor docs | yes: review the diff, open a pull request, merge Conductor docs | yes, the Free plan brings your own subscriptions Conductor pricing | local macOS app; cloud workspaces on paid plans Conductor installation docs |
| OpenHands Agent Canvas | yes, one worktree per agent Agent Canvas page | yes, Claude Code, Codex, Gemini CLI and ACP agents Agent Canvas page | partial: manual compare only, no per-task judge OpenHands ACP post | not documented Agent Canvas page | yes, “use your existing subscription” Agent Canvas page | local agent server, or OpenHands Cloud OpenHands docs |
| Blackbox AI /multi-agent | yes, each agent on its own git worktree Blackbox CLI page | yes, Blackbox, Claude Code, Codex, Gemini Blackbox CLI page | yes: a Chairman LLM evaluates and selects; you can override or merge Blackbox CLI page | yes, the chosen result lands as a scoped pull request Blackbox agents page | not documented; enterprise agreement and credits Blackbox agents page | CLI, and an HTTP service Blackbox agents page |
The closest match
Blackbox AI has the same shape. Four things differ.
Blackbox sends one task to several vendors’ agents, has a Chairman LLM pick, and lands the pick as a pull request. Judging by itself is therefore not what sets Cyberine apart. What the Blackbox pages do not describe:
Adversarial skeptics
AdversarialJudge sends several skeptics at each candidate, each prompted to refute it rather than approve it.
Your seats, not credits
Claude, Codex and the other CLI agents run on the logins you already hold, and Cyberine never reads the credentials. How cost is kept in check.
Local execution
Everything runs on your machine, in your repository. There is no hosted service between you and the agents.
Limits of this comparison: Blackbox’s selection criteria come from its two product pages only, since its docs pages could not be read. Bringing your own subscription is not exclusive to Cyberine: Conductor and OpenHands do it too. Cyberine has no cloud execution, which Cursor, Codex, GitHub, Conductor, OpenHands and Blackbox all offer.
Run your first race.
Install it, point it at a repo, give it a task.
run 81cbc07cangle: the fieldrecorded 2026-09-254 workers3 diffs applied1 kept$0.10 spent