Runway

Most agent runs leave you with a diff and a hope. Orbit structures that run into what the vendor calls an ‘orbit’: the harness selects one task from a dependency-ordered backlog, hands it to whatever coding agent you’ve configured, runs your tests, lint, and type checks as validation gates, and only closes the task if the agent can prove the result. Every run writes four artifacts — a structured agent-result.json, a rubric-scored evaluation.json, a review.json with an accept/iterate/stop recommendation, and a human-readable progress.md — so the evidence trail is inspectable long after the run completes.

The differentiating feature is the gate-and-artifact contract, not the agent. Orbit is described on the vendor page as agent-neutral: Claude, Codex, Cursor, or any CLI that speaks JSON can be dropped behind the same harness. That means you compare agents by inspecting their artifacts under identical conditions instead of relying on anecdotal impressions from different sessions. The vendor explicitly calls out ‘adapter experiments’ as a supported workflow — swap the agent, replay the same task, diff the evaluation scores.

Orbit fits narrowly: repositories with real automated test suites, teams that need reproducible audit trails for agent behavior, and harness engineers prototyping how agents handle a backlog. It does not supply the agent, the tasks, or the test suite. Teams that have none of those pieces in place will be building the prerequisites before Orbit adds value. The vendor page describes the project as ‘intentionally small,’ which is an honest signal about scope — contributions are welcomed specifically when they make the harness easier to verify or replay, not when they expand what the harness does.

The deterministic replay demo — invoked with MOCK=1 ./replay.sh auth-rescue — requires no API key and demonstrates the full selection, validation, and artifact-recording loop locally. The project is MIT licensed with no paid tier described on the vendor page.