Insights
Public notes on the decisions that appear before the code — ADRs, runbooks, analyses, and issues of the newsletter “Architecture in the AI Era”.
Newsletter
A monthly newsletter for architects, EA leads, and CIOs. Five issues published. Each one anchored in a real project or field evidence — never in a market trend. The subject is always the same slice of the problem: the architecture, identity, data, and control that decide whether an AI pilot reaches production or dies in the demo.
The editorial rule is simple: no issue ships without a concrete case behind it. When the material comes from client work, the content is anonymised and sensitive data is redacted before publication.
Issue 1
The founding thesis: the model is 5% of an AI project; the other 95% is the architecture underneath it — integration, identity, data, and governance. A reframe follows from that: your AI strategy is your architecture strategy.
Published in English · opening issue
Issue 2
Agents in production without a distinct identity are unmanaged service accounts with a chat interface. 63% of enterprises cannot limit an agent's purpose after it ships; 60% have no kill switch. Enterprise identity infrastructure was designed for humans, and agents do not fit inside it.
Published in English · field evidence on agent identity
Issue 3
One weekend, one Arch Linux desktop, one agent with shell access. The agent audited its own access, reported the permissions that had never been explicitly revoked, and found a second default credential that a human reviewer usually misses. The shape of the enterprise problem, compressed into a single machine.
Published in English · controlled experiment on own hardware
Issue 4
Ten years of HR data archived from Oracle to Databricks. The questions that stalled the project were not about the pipeline — they were about governance: where to redact, who can see unmasked data, and how to prove control. AI readiness does not start with the catalogue; it starts with who can see what.
Client work — anonymised and redacted
Published in English · real project, data and client anonymised
Issue 5
Two AI gateways announced in the same month: one shipped with the controls this newsletter has been asking for since issue 1 (kill switch, audit trail, containment), the other with no public evidence that those controls exist. The difference is not the model — it is the control architecture around it.
Published in English · public announcement read against audit criteria
The newsletter is distributed by email and archived privately, so there is no public permalink per issue yet. Rather than link to a page that does not exist, these cards state the thesis and the evidence behind each issue. When an issue gets a public page, the card becomes a link.
To receive the next issue, write to
contato@arquitechcorp.com
with subscribe in the subject line.
Beyond the newsletter
Not every note becomes an issue. The posts below are what circulated on the professional profile — a legacy integration case, a containment test, and a warning about agent sprawl.
A VM from 2010, 80 GB of logs, an Azure Storage Account, and a cryptographic handshake that could not agree on an algorithm. The way out was not to force the old VM to speak modern — it was to place the control point where the team has authority.
55% cannot isolate an agent from the network if something goes wrong. Does your agent have a kill switch, and was it tested this month?
AI agent governance is not about the model — it is about how many agents you have, who created each one, and whether you can list them all today.
The posts and the discussion thread for each issue live on the professional profile. Messages and comments are welcome, and the direct channel for a technical subject remains email.
Public method
The architecture decisions behind ArquitechCorp's own operation are recorded as ADRs (Architecture Decision Records) and maintained as portfolio assets. The method is public; sensitive content is redacted before publication. This site is the first asset built under that discipline.
The value of a decision record is in the context and the reasoning — not in the client's data. So what gets published is the reasoning, and what gets removed is anything that identifies third parties.
Thirty minutes, no institutional presentation. Bring the symptom — integration degradation, a cloud bill, a stalled AI pilot — and we leave the conversation with the initial technical hypothesis and what would need to be measured to confirm it.
No form in this first version: a form here would mean processing personal data without a declared legal basis. Direct channel, with a reply within one business day.