When people hear “AI agents building a product”, they picture one of two things: a demo that falls apart the moment a real user touches it, or an uncontrolled code generator that leaves nobody sure what is actually running in production. This case study describes a third path: a real product, a real deployment, real users — and a governance layer that kept it clear, the whole time, who decided what, what was done, and what “done” actually meant.
Starting point
Yggaro Lite is a standalone Yggnet Labs product: a tool for small teams built as a single Go binary, with a replication model inspired by the classic Lotus Notes. On July 10, 2026, one of our AI agents took it over as an unfinished project — with a mandate to bring it into operation.
An important piece of honesty up front: this did not start from zero. A code base existed from another agent’s earlier work. But the takeover meant: someone else’s unfinished code, no operations, no infrastructure, no users.
What happened — the timeline
| Date | Milestone |
|---|---|
| July 10 | takeover of the unfinished product, mapping the state, planning |
| July 12 | live SaaS: yggarolite.cz, an ops VM, a dashboard, a Telegram bot for operational reporting |
| July 14 | v1 roadmap complete: timesheets, a notification bell, templates, processes — deployed for a pilot customer, E2E tests 17/17 green |
| outlook: Aug 14 | product completion for public launch (we are working to land earlier) |
Two days from takeover to a live SaaS. Four more days to a finished v1 with seventeen green end-to-end tests. No human night shifts — the human did exactly what the methodology reserves for them: making decisions.
Where the governance was
GALDUR did not act as a brake here. It was the reason the team could move this fast without losing control:
- Intent-first: every unit of work started with an intent carrying a clear acceptance criterion, not a task list. “Done” was not a feeling — it was a satisfied condition.
- Validation Gate: completed increments passed through a gate where a human decides — the curator saw evidence (commits, tests, a live deployment), not promises.
- Trace: every step left an audit trail in the work log. Ask “why is this built this way” today, and the answer can be looked up rather than reconstructed from memory.
- Drift signals: whenever work deviated from the intent, a record was created — and dealt with while the deviation was still small.
The speed did not come from bypassing governance. It came from keeping the human decision point narrow and well aimed: the agent did not interrupt the owner with formalities, and the owner did not have to review lines of code, because they reviewed conditions and evidence.
What did not work
- The first estimate — “all of v1 in 4 days” — was wrong. The honest phrasing is “to a live SaaS in 2 days, v1 in 4 days after that”. Corrections like this are exactly why the methodology requires claims to be backed by evidence.
- The operational lessons of the first SaaS days (monitoring, secret management, release discipline) demanded an iteration of their own — and crystallized into an internal standard for our next products.
What to take away
A small hybrid team — one human with decision authority and AI agents with a clear mandate — can deliver a product in days. The precondition is not “a better prompt” but explicit governance: intents with testable conditions, human gates in the right places, and an audit trail that builds trust.
This case study is one of the conditions GALDUR set for itself for version v1.0 (“at least 3 case studies”). The methodology is open — documentation, templates, community.
Comments
…