Educating the agent, not launching it: the Azure session of August 29
In short: what matters is not having succeeded in deploying the Tower on Azure. What matters is that we wrote down what blocked us, and how we unblocked it. A full day went into writing the deployment circuit — every error recorded, so that next time we do not go through the same traps again.
The day, without selling it as a success
We spent the day trying to connect the Tower to Azure. Not for the pleasure of the cloud: so that the Tower's agents can talk to Azure the way they already talk to their living map.
We did not finish. That is important to say first.
But we did something more valuable: we wrote the circuit. Every error, every blockage, every solution found together was recorded in circuit-deploiement-alice-cloud.md. Not to make it pretty. So that tomorrow, another agent — or the same one, or yourself — can replay the same sequence without going through the same traps.
The agent that says "I am stuck"
Today, at one point, the agent stayed stuck. Not because it did not know what to do, but because it was missing information I had forgotten to give it: the Azure tenant did not exist, the API key was not working, the GitHub token did not have the right permissions.
The agent told me where it was stuck. I gave it what was missing. And off we went again.
What is simple is this: the agent explains what is missing, the human provides it, and the circuit gets written. No mystery, no crisis. Just a structured conversation between an agent that learns and a human that decides.
What we unblocked today
The Docker image of the Tower Community is published on ghcr.io/fotsopatrick/tour-community:latest (build succeeded, 5a29035).
The Azure AI Foundry keys are valid (HTTP 200 on the model catalogue).
The deployment circuit is ready, with the variables that will be needed once the Azure tenant is functional.
What we learned
Three lessons, taken from the day:
- Regional quotas on Azure are a non-event if one uses GlobalStandard — which is what we did.
- The deployment name must be exactly the model's name (gpt-4.1-mini). A name mismatch is not reported by Azure: it shows up as a slow error.
- The history of failed commands is worth more than the successful ones: they tell what not to do again.
What remains to be written in the circuit
Not for tomorrow, not for the day after. For next time.
The following blocks are noted in the circuit, waiting for a real Azure identity (az login), a working Azure OpenAI deployment, and a real GCP tenant. They are not erased. They are written. When we have those elements, we will complete the circuit and it will be replayable from start to finish, without error.
This is a wall, not a promise
A wall is not a perfect tool that does everything on the first try. It is a structure that resists errors, records them, explains them.
The Tower does not promise to succeed on the first try. It promises that every failure becomes a lesson, and that every lesson becomes a procedure that all agents can read.
Today we wrote one page of the wall. Tomorrow we will write another.