Operating model
Intent over manual translation
The goal is not to remove expertise. The goal is to stop spending expert time translating the same intent into repetitive configuration.
Declared outcomes
Describe what done looks like, not the steps to get there. Agents should discover the path within constraints.
Compounding context
Better context beats a bigger model. Curate architectural facts, standards, and constraints agents can reuse.
Memory before autonomy
An agent should not act on a cloud estate it does not understand. Inventory, ownership, history, decisions, and exceptions come first.
Trust and safety
Guardrails before autonomy
Policy, budget limits, blast radius, and approval gates come before agent action.
Reversibility
Speed is safe when rollback is cheap. Treat irreversible actions as gated exceptions.
Explainability
If an agent cannot explain why it built something that way, it owns your cloud, not you.
No silent production changes
Production changes require policy, evidence, approval, rollback, and a record of what changed.
Least privilege agents
Agents should start read-only, earn permission through trust, and act only inside scoped identities and blast-radius limits.
Architecture
Portable intent
If your description only works in one provider, it is config, not intent.
Conway's cloud
Your infrastructure mirrors how you describe it. Clear intent creates cleaner systems.
Human in the loop
Automate the toil. Keep humans on risk, taste, tradeoffs, and judgment.
Audit by default
Every recommendation, approval, action, exception, and rollback should become part of the operating record.