// AI & Automation — 03
Automation with receipts.
Event-driven operations, LLM integrations, and agentic tooling for business systems — built with the same engineering discipline as everything else: explicit boundaries, auditable behavior, and honesty about what AI is actually good at.
// A — Principles
How intelligent automation is scoped here.
01Automation first, intelligence second
Most operational wins are plain automation — triggers, queues, retries. LLM steps are added where reasoning genuinely pays, never as decoration.
02The queue is the product
Idempotency keys, retry budgets, and a dead-letter path. An automation you can't pause, replay, and audit is a liability with a nice API.
03Guardrails and evals or it doesn't ship
LLM steps get defined inputs and outputs, evaluation sets, and cost caps. 'It seems good' is not a deployment criterion.
04Human gates where consequences live
Drafting, classifying, extracting — automate freely. Anything that spends money or contacts customers keeps a human review gate.
05Audit everything
Every automated action is logged against the record it touched. When someone asks 'why did the system do that?', the answer is a query.
06Agents as executors, not deciders
The useful pattern is a tireless executor following a structured runbook — with the verification steps it's tempted to skip made mandatory.
// B — Architecture
The shape of the system.
FIG — THE AUTOMATION PATTERN. TRIGGERS IN, GUARDED REASONING IN THE MIDDLE, AUDITED ACTIONS OUT.
// C — In practice
Where this runs today.
Production automation in messaging, reminders, and event processing — plus open experiments on the agentic edge.
// Engage
An operation that should run itself?
Reminder pipelines, webhook processing, LLM-assisted ops with real guardrails — scoped and built end to end.