Six articles that read the public code and the legal texts behind an agent platform: the stack, its vendor-free variant, what the regulations ask, how Microsoft's governance toolkit enforces it, what gets recorded, and how a model is rolled out without breaking any of that. Read in order; each part builds on the one before.
| Part | Article | Reading | Link |
|---|---|---|---|
| Part 1 | AI agents on OpenShift: the open blueprint, layer by layerRed Hat's July 2026 reference architecture read against the code: two exits per agent, two execution patterns, the two gaps Red Hat names itself. | 6 min | Open |
| Part 2 | The same agentic stack with no platform vendorEvery Red Hat box replaced by the CNCF, NVIDIA or Microsoft project it packages; what "free" means; what you take on. | 7 min | Open |
| Part 3 | What the regulations actually ask of an agent stackEU AI Act as amended in July 2026, NIS2, ISO 42001, Swiss FADP, FINMA 08/2024, mapped to the layer that has to answer. Not legal advice. | 9 min | Open |
| Part 4 | Action governance with Microsoft's Agent Governance ToolkitPolicy per action, declared intent, fail-closed approval, Merkle audit; three insertion points; the toolkit's own manifests and sector templates. | 11 min | Open |
| Part 5 | Observability: what the stack records, and what it cannot tell youFour records for one tool call, what an auditor asks, the six-month retention problem, why a Merkle tree needs a witness. | 7 min | Open |
| Part 6 | Rolling out a model blue/greenTwo llm-d pools behind one HTTPRoute, an Argo Rollouts analysis as the gate, garak probes, a person who promotes and is recorded. | 9 min | Open |