ultrastableCoding agents write code. Ultrastable runs the system.
It finds its own work in your metrics and failures, hands each task to the coding agent that fits, proves every change live with an automatic revert, and keeps score of how often its decisions were right.
7 of 7 claims pass their live check right now (checked 2026-11-11 09:12 UTC).
What is different
It chooses the work
It proves every change
It keeps score
Works with the agents you already use
Ultrastable is the layer around coding agents, not another one. Open models on its own code loop, GitHub Copilot and Claude Code sessions all take work through one claim ledger, so two agents never edit the same file.
Where it fits
| Approach | Who chooses the work | Who checks it |
|---|---|---|
| Vibe coding builders | A person, prompt by prompt | The person, if at all |
| Coding assistants (Cursor, Claude Code, Codex CLI) | A person; the agent multiplies effort | The person reviews |
| Cloud coding agents (Devin, Codex cloud, Copilot agent) | A person assigns a task | CI and a reviewer |
| Agent governance and gateways | Not their job | Policy, cost and access logs |
| Ultrastable | The system, from its own metrics and failures | Live probes, automatic revert, a two-family review and scored predictions |
Proven in production
It has run a live system since August 2026: thousands of merged pull requests, a hundreds-strong set of failure detectors and a scored prediction history. It is one installation, and the honest gaps are listed on this site.