Enforra

Orbit frames itself as mission control for AI coding agents, not an agent itself. The core loop is: select a dependency-ordered task from a backlog, hand it to a configured agent adapter, run the validation suite — tests, lint, type checks — and record the outcome as durable JSON artifacts regardless of whether the orbit closes or fails. Nothing advances until validation passes. That gate is the architecture, not a setting.

The differentiating feature is the artifact contract. Every run produces four files: `agent-result.json` with structured output and changed files, `evaluation.json` with rubric scoring across task focus and diff signal, `review.json` with an accept-iterate-stop recommendation, and `progress.md` as a human-readable mission log. This means you are comparing agent outputs against identical evidence rather than against your memory of what the last run looked like — the docs describe this explicitly as enabling adapter experiments, swapping coding agents behind the same contract.

Orbit fits teams running multiple coding agents who need an answer to ‘which agent actually solved the task?’ — the artifact replay demo runs without an API key, which the vendor positions as a signal that the harness is intentionally verifiable independent of any one provider. It does not fit teams that need autonomous backlog management, goal decomposition, or any planning layer above the task level. Those capabilities live in the agents Orbit wraps, not in Orbit itself. The codebase is MIT licensed with no commercial tier — what ships on GitHub is the entire offering.

The self-hosted path requires Python and a virtual environment; the replay demo uses a MOCK flag so teams can inspect the full artifact pipeline before connecting any live agent. Orbit describes itself as intentionally small, and community contributions are scoped to adapters, demo missions, and templates — not new planning abstractions.