CoreTex

Orbit positions itself as mission control for AI coding agents: it takes a dependency-ordered backlog, selects one task per run, routes it through a connected coding agent, then validates the result against tests, lint, and type checks before closing the loop. The core workflow is deterministic and replayable — the vendor provides a demo that requires no API key and walks through task selection, agent execution, validation, and artifact recording as a single inspectable sequence. Every orbit produces four artifacts: agent-result.json (structured output and changed files), evaluation.json (rubric scores for task focus, completion, and diff signal), review.json (accept, iterate, or stop recommendation), and progress.md (a human-readable mission log). None of that evidence disappears between sessions.

The differentiating feature is the validation gate. The orbit does not close unless the agent can prove completion — if tests fail, the loop does not advance. That single constraint is what separates Orbit from agents that self-report success and move on. The rubric scoring in evaluation.json adds a second layer: it surfaces whether the agent stayed focused on the task, produced meaningful diff signal, and met validation, which makes swapping agents an objective comparison rather than a judgment call.

Orbit fits teams running one or a few coding agents on the same repo where test and lint discipline is non-negotiable, and where a durable audit trail matters — regulated environments, open-source projects with strict contribution standards, or any context where ‘the agent said it worked’ is not an acceptable handoff. It breaks down when you need agents working in parallel on separate tasks, when your workflow requires branching based on what a prior step returned, or when you want to connect Orbit directly into a CI pipeline without extending the harness yourself. The vendor’s own framing — ‘intentionally small’ — is accurate and should be taken literally.