Practical notes from teams shipping AI agents in production — design patterns, audit-ready architectures, and the operations playbook we wish we'd had when we started.
Most writing about AI automation is either vendor copy or research abstraction. This is neither. We publish what we learn running real workflows for operations teams: where AI agents genuinely replace manual work, where they need a human checkpoint, and what breaks in production once volume arrives.
Recurring themes: human-in-the-loop design — when to route an agent's action to a person and how to make that approval take five seconds rather than five minutes; audit-ready architecture — why every run should produce a tamper-evident trail before you ever need it for SOC 2 or a customer dispute; document and invoice processing — extraction accuracy, confidence thresholds, duplicate detection, and posting into Xero or an ERP; and the Singapore operating context — PDPA alignment, GST handling, and the PSG and EDG grants that fund a large share of local automation projects.
Posts are written for the operator, not the integration engineer. If a piece can't be acted on by the person who owns the process, we don't publish it.