Every guide covers the same 10 chapters — locker, trigger, prompt, tools, verify, output, guardrails, observability, cost, and a one-line deploy — specialized for the failure modes of that loop category.
Cron-driven Claude agents that read the internet, diff yesterday, and post a digest your team actually opens.
Autonomous PR-opening agents that pass CI on the first try — because they run CI locally before pushing.
Multi-step Claude agents with tools, memory, and a supervisor that says 'stop' before the bill does.
Ingest → chunk → embed → retrieve → re-rank → answer with citations — and an eval that catches drift.
Webhook-triggered Claude agents that triage alerts, correlate signals, and page a human only when it's real.
Ticket in → Claude drafts a reply grounded in your docs + prior tickets → agent one-clicks send.
Enrich the lead, personalize the opener, log the activity, hand a warm reply to the AE.
Nightly reconciliation, anomaly detection, a board-ready daily flash with numbers Claude did not hallucinate.
Continuous secret + dependency + IaC scans, deduped and triaged by Claude, with SOC2 evidence auto-collected.
Meeting notes, weekly rollups, and inbox triage — quiet loops that give people back an hour a day.