How do you ship RAG that survives production traffic?

Retrieval is treated as a data product: connectors, chunking and metadata, hybrid search where ROI is clear, staleness detection, and answers constrained to cited context. We gate rollout behind scope limits, policy prompts, and PII redaction.

Where do you draw the line between agents and deterministic workflows?

Ambiguous work gets tool augmented agent loops with budgets and timeouts; money moving and SLA critical steps live in explicit graphs with typed I/O and idempotency keys. HITL gates prevent silent changes to authoritative state.

Which n8n patterns do you standardize for scale?

Composable sub workflows, dedicated error and DLQ paths, queue mode when throughput requires it, secrets via your vault, execution records suitable for replay, and alerting on failure rate, p95 latency, and queue depth.

How do you keep LLM calls inside n8n maintainable?

Orchestration stays in the graph; model policy lives in versioned prompts, structured outputs, and explicit schemas for tools and HTTP steps. Streaming when UX demands it; batching otherwise for simpler retries and cost attribution.

What does observability and cost control look like end to end?

Correlated request IDs across workflows and model calls, token and latency dashboards by route, redacted failure payloads for replays, concurrency caps aligned with provider limits, and staged rollouts when changes touch customer visible behavior.

Do you build operator software on top of automations?

Yes—review queues, approvals, admin consoles, and external portals so teams override, annotate, and replay runs without touching the canvas. Those surfaces often close the loop into eval labels and improved retrieval or prompt versions.