Patina

Most AI coding agent runs are fire-and-forget: the agent edits files, you check the diff, you guess whether it worked. Orbit replaces that guess with a structured loop. The harness selects a task from a dependency-ordered backlog, hands it to the agent, runs real validation (tests, lint, type checks), and only closes the orbit if the checks pass. Every run produces 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. The evidence trail is there whether you review it immediately or audit it a week later.

The differentiating constraint is the validation gate. The agent cannot self-report success. It has to prove completion through the checks you configure — failing tests, lint violations, and type errors are evidence the loop uses to decide whether work advances. The vendor describes this as ‘if the agent cannot prove it, the orbit does not close.’ That design makes Orbit useful for teams who have been burned by agents that return confident output on work that does not actually pass a real check.

Orbit is agent-neutral by contract: the docs describe support for Claude, Codex, Cursor, or any JSON-speaking CLI, and the contribution model is centered on swappable adapters. This makes it a practical bench for comparing agents on the same task with structured, comparable artifacts rather than subjective impressions. The tool is MIT licensed and self-hosted only — there is no managed runtime, no hosted scheduling layer, and no external platform dependency. A deterministic replay demo requires no API key and runs locally in minutes.

Where Orbit breaks: the harness is scoped to one task, one loop, one agent at a time. Teams that need agents running in parallel, cross-agent handoffs, or a hosted execution layer will hit the ceiling of what Orbit is designed to do. The project’s own documentation frames it as ‘intentionally small,’ and contributions are explicitly invited to make it ‘easier to verify, easier to replay, or easier to connect to another coding-agent workflow’ — not to expand it into a full orchestration platform.