SIMD Agent

Most AI coding agent demos end at the diff. Orbit picks up where they stop: after the agent returns output, the harness runs your actual test suite, lint, and type checks — and if validation fails, the orbit does not close. The core workflow is a loop: Orbit reads a dependency-ordered backlog, selects the next eligible task, hands it to whichever CLI agent is configured, collects structured output, runs the configured validation gates, writes four artifact files, and either advances or flags for human review. The deterministic replay demo — requiring no API key — lets you inspect this entire cycle locally before connecting a live agent.

The differentiating feature is the artifact layer. Every orbit produces `agent-result.json` (structured status, changed files, raw agent output), `evaluation.json` (rubric scoring across task focus, completion, diff signal, and validation), `review.json` (accept, iterate, or stop recommendation), and `progress.md` (a human-readable mission log). This means compliance reviews, retrospectives, and agent comparisons work from evidence rather than memory — you can diff two agents’ evaluation scores on the same task contract without trusting anyone’s anecdote.

Orbit fits tightest in codebases that already have a working test suite and lint configuration, because the validation gates are only as strong as the checks you wire in. It fits least in projects that need a hosted interface, webhook integrations, or multi-repo coordination — the vendor page describes the project as intentionally small, and the contribution guide explicitly targets contributors who want to make the harness easier to verify or replay, not to grow its surface area. Teams that outgrow the CLI harness model — because stakeholders need a browser dashboard or because the backlog spans more than one repository — are the ones who move to a purpose-built agent platform.