Coding assistants write the code you ask for. Ultrastable decides what to build next, ships it, proves it live, and keeps score of how often it was right.
The problem
I run QNFO, an independent research operation, largely through AI agents. Agents can write code quickly, but nobody was deciding what they should work on, checking that a change worked once it was live, or noticing when the system's own judgment was drifting. That is a systems problem, not a coding problem: the gap is in goals, feedback and accountability.
What the platform does, under my direction
- Its own backlog. The platform reads its metrics, failures and probes and files its own work as issues, each with a closing probe.
- Several coding agents, one set of rules. It hands each change to coding agents from different vendors under one written agreement and one claim ledger, so two agents do not edit the same thing.
- Proof, not promises. Every deploy is checked against live probes, and a change that breaks one is rolled back automatically.
- Scored judgment. Before it acts, it logs what it expects to happen; afterwards it checks the outcome and keeps a trust score per kind of action.
- Nothing silently lost. Every issue is open, closed with evidence, or a recorded delete.
I set its goals, standing directives and limits (what it may decide alone and what stays with a person) and I evaluate how it behaves. That is the same work I did as a network modeler and at the U.S. Federal Highway Administration, where I specified a national travel model and directed the contractor team that built it.
Checked live, right now
These claims are re-checked by the running system against its own database. Loading the current results from ultrastable.io…
549 issues with a code task filed in the last 7 days (2026-11-10).
774 of 1,059 scored decisions matched the prediction logged first (2026-11-10).
6,041 issues closed with a disposition, 0 silent drops (2026-11-10).
634 failure modes recorded, 611 active detectors (2026-11-10).
Also: 3,880 merged changes between August and November 2026, 624 of them opened by the platform's own code loop.
What running it taught me about evaluation
Proxy quality signals (test pass rates, reviewer ratings, issue counts) can look healthy while a system quietly drifts. The drift showed only when the platform's decisions were scored against the predictions it had logged first, so its quality is now judged by outcomes, not by proxy scores. The capability I think benchmarks miss is the one this exposes: whether a system can be trusted over time in live operation, catching and correcting its own errors.
What it does not do yet
- It has one installation: QNFO. A second, independent installation is the next milestone.
- Hundreds of issues are open against a target of zero.
- Its decision scoring is graded by the system itself; outside evaluation is still to come.
Why this matters beyond QNFO
Teams adopting coding agents face the same questions: who decides what the agents do, how do you know a change worked, and how do you audit an AI system that acts on its own? Ultrastable is my working answer, and the evidence above is how I would want any autonomous system held to account. I am open to roles in AI governance and oversight, AI evaluation, autonomous systems strategy and research leadership, and to conversations with teams that want to run their own work this way.
Contact: rowan.quni@qnfo.org · LinkedIn