Not a promise. Proof: a platform built end-to-end by AI agents

Our website carries a sentence that deserves suspicion: “A complete real-estate platform in production — built end-to-end with Coautra.” Anyone can claim that; in 2026, every second agency seems to. So this post lays open what is behind it — in five sections, with real numbers, and including the places where it creaked.

1 · The starting point

Immotron is a product of UniPlacement GmbH: a real-estate platform with a web presence, a backend API and everything that belongs to it. And it is our test case for a conviction I have carried around for ten years: that the complete development loop can be closed — from vision to user stories to issues to code, tests, documentation and release, without knowledge getting lost between any two steps.

Coautra is the name of that process. Not a tool you buy — an audit-proof workflow in which AI agents do the work and every step is captured, reviewed and documented.

2 · What the loop looks like in practice

A requirement — often literally dictated as a voice message — always passes through the same chain: it becomes a structured user story, the story becomes issues with acceptance criteria, the issues go to development agents. Several of them work in parallel, each in its own isolated working copy of the code. What they deliver never lands directly on the main branch: every change becomes a merge request, runs through automated tests and quality gates — end-to-end tests included — and the documentation, manual included, is updated in the same pass, not “later”.

At the end stands a release. And at the start of the chain stands a human with a phone.

3 · The numbers

As of this week, pulled straight from the git system, core repositories only:

over 1,700 merged merge requests (API: 682 · web: 669 · the process system itself: 387)
over 2,000 issues, each born from a requirement, most of them worked off by agents
hundreds of isolated working copies — one per task for most tasks, so that agents working in parallel never get in each other’s way
one production system that is running while this text goes live

Numbers don’t prove quality — but they prove operation. This is not a weekend prototype and not a pitch-deck demo.

4 · What went wrong (the most important section)

Honest field reports owe you this part. Three examples from the actual history:

Agents deliver plausible nonsense when the requirement has gaps. The process doesn’t catch that with better models, but with acceptance criteria in the issue and tests that check exactly those criteria. Without both, multi-agent development is an accelerant for chaos.
Escalation is a feature. There are tasks where the agent must stop and ask a human — budget limits, architecture decisions, ambiguity. The art is not avoiding escalation but making it cheap: one question via Telegram instead of a failed sprint.
The process itself is 387 merge requests old. The system that orchestrates the work was built and rebuilt just as iteratively as the product — and several of those rebuilds were corrections of its own wrong decisions, traceable in the decision logs.

5 · What this means for an SME

The point of this proof is not “AI can code”. The point is: the loop can be closed when the foundations are right — requirements as structured text, systems with APIs, tests as a contract, documentation as part of the definition of done. Those are exactly the same foundations that make a company AI-ready, even if it never wants to build a platform of its own.

And the legacy systems? On this road they become the companion: the same process that builds new things also documents and modernizes the grown system — controlled and traceable instead of one risky big-bang.

If you want to think this through on a concrete case: write me — a conversation, not a sales funnel.