Andre Semek Cloud & AI Solution Architect, Berlin

I design cloud platforms. And systems where AI agents can do real work without me having to trust them blindly.

Solution design

In my job at a large enterprise I design platform solutions on Azure. That covers AKS, infrastructure as code with Bicep and Terraform, and delivery through Azure DevOps and ArgoCD.

Platform and DevOps

For seven years I ran platforms and moved systems from Docker Swarm to Kubernetes, on GCP, AWS and Open Telekom Cloud. With Terraform and Ansible, secrets in Vault and monitoring through Prometheus and Grafana.

AI in engineering

At work I built an agent workflow from story to pull request. At home I run my own agent platform with routing, a review gate and guard hooks.

Trust has to be built in

An agent that writes code also runs commands. So the rules do not sit in the prompt but in the path every change has to take.

  1. TaskThe goal, the files it may touch and the command I check the result with.
  2. BrokerPicks model and tool by cost, quota and earlier failures.
  3. AgentWorks in a sandbox with no keys and no root.
  4. HooksStop dangerous commands before they run.
  5. ReviewTwo models from two vendors. A BLOCK holds the change.
  6. mainOnly now is it merged. Publishing needs my own go-ahead.
How a change gets from a task to main in my own agent platform.

Limit: the review environments still run without an OS-level sandbox. I still have to take care of that.

$ ops list
git       commit, push, open and merge a PR through the review gate
review    send a commit to the review gate and follow the findings to a fix
council   decide with a council: web research first, then three members debate …
handoff   update the HANDOFF.md top section from this session and verify its paths
radar     decide whether a tool, repo or model is worth adopting, and document the verdict
…
An excerpt from my list of workflows. Recurring work has fixed names, for example for review, decisions and handoffs.
All the guardrails

Decisions and what they cost

All four cases
  1. All tests green, the watcher was dead

    context-parachute saves the work state before an agent’s context window fills up. After a change the watcher stopped running, with no error message. The tests still all passed.

  2. My router did worse than a fixed order

    The router ranked accounts by success rate. New accounts with no history therefore always lost to proven ones, and paid quota went unused.

  3. A worker used sudo

    A delegated worker installed a system package with sudo because it was missing a tool.

Journal

I write down what I learn from working with agents: how much context a task needs, what a team setup takes, where the costs come from.

Write to me if you want to talk about any of this.