The Platform That Builds Itself - Opengeni
Opengeni
SubscribeSign in
The Platform That Builds Itself<br>Welcome to the Opengeni blog, picking up the story we started on the Cloudgeni blog, several months and one strange loop later.
Opengeni<br>Aug 12, 2026
Share
Welcome. Back in May we published Introducing Opengeni: An Agentic Runtime for Organizations. About open-sourced the platform we had spent two years building, the engine that lets AI agents do real, long-running work while humans stay in charge.<br>This post picks the story up a few months later. And the most interesting thing that happened in that time was not new features.
We became our own heaviest user
Somewhere in early summer, the way we build Opengeni changed. We didn’t just built the platform but started running on it. Our own agents, pointed at the platform they run on.<br>It starts simple. You have a runtime built around one idea: give an agent a goal, real success criteria rather than a prompt, and let it keep working until the job is actually done. Some of these runs can stretch over days. A restart does not cost it its train of thought, and it stops to ask a human before doing anything consequential. So one evening you hand it a goal on your own codebase, on your own build server, and you go home.<br>Next morning there is a pull request waiting. Someone reads it, asks for one change, approves. It merges.<br>Then you do it again. And again. At some point it dawns on you that the loop has closed: the platform is both the worker and the thing being worked on. An agent running on Opengeni fixes a rough edge in Opengeni, and sometimes that rough edge is one the same agent tripped over the night before. Every improvement makes the next run a bit more capable, so the next improvement lands a bit sooner. A real flywheel effect.<br>Some numbers from that flywheel:<br>Since the May release, nearly 2,000 commits have landed on our main branch or 1300 Pull Requests. We are a team you can count on one hand, and nobody is crunching. The pace comes from agents grinding through the long, patient work overnight while our mornings go to the parts that actually need a person: review, judgment, direction. The amount of work is similar that we produced after working almost two years on Cloudgeni prior to that.<br>One agent spawned 1000+ subagents to do the work. Now, admittedly, perhaps not all of that was nececery, and we had to cut down on some stuff. But genuinely, if you count the amount of work, planning, architecting, creating issues, consolidating, running tests, doing deployments to different environments, testing… you can understand where it comes from. There is genuine need for agents delegating this work and it was amazing to see that runtime handled it without a hitch.
What kept it from being terrifying
Now for the part we care about most, because it separates a good story from a reckless one: at no point did we not have full control as well.<br>Every run works toward an explicit goal and cannot simply wander off and declare victory. Finishing takes evidence. A run that stops making progress gets paused on its own. Anything consequential waits for a human eyes. Everything the agent does, down to each step and approval, lands in a permanent record we can replay later, so the question of what it did while we slept stops being a black box and becomes a search query. Agents have limited permissions that we gave it, they run in isolated sandboxes and can do only what we want them to.<br>Put differently: the guardrails we originally built for platform teams turned out to be the very thing that let us trust the loop.<br>We didn’t let go of the rules to go faster. The rules are why we could go faster.
What changed on the platform, briefly
The proper tour can wait for the next post, but in broad strokes, since May: agents can now run directly on your own computers (a laptop, a build server), not just in cloud sandboxes, and nothing on your side has to be exposed to the internet. They remember what your workspace teaches them, though a human signs off on every memory first. They meet you where you work (in Slack, even on a voice call) rather than only in a dashboard. And they have gotten better at the thing this whole story began with: taking a goal and finishing it.<br>All of it is open source, Apache-2.0, at github.com/Cloudgeni-ai/opengeni.<br>What this blog will be
Expect two kinds of posts. Agent Sessions : essays on agentic development at large (why agents stop too early, what production really demands, what running agents on live infrastructure keeps teaching us). And Opengeni Notes : when something ships that is worth your attention, we will say so and explain the lesson behind it. The next post is one of those, a fuller tour of the summer for the technically curious.<br>If this sounds like your kind of problem, subscribe. And if you would rather read code than blog posts, fair enough: the repository is open and so are the issues.<br>Built in Oslo, Norway. Increasingly, by the...