Topic: #Engineering

How a multi-tenant sales agent system grew from a Google ADK hello-world into a 40-worker orchestrator, then got rebuilt on ADK 2.0 Workflow graphs: what broke, what held up, and what we'd tell any team making the same jump.

Nested human-in-the-loop workflows fail intermittently when an interrupt matches the wrong sub-workflow, or when a resume reads state before the pause write lands. How ADK 2.0, LangGraph, LangChain, CrewAI, AutoGen, and n8n handle both halves.

A multi-agent coordinator that re-classifies every message will restart mid-flow workflows the moment a human pauses them. The same fix — an active-workflow marker checked before routing — across ADK 2.0, LangGraph, LangChain, CrewAI, AutoGen, and n8n.

We shipped a backend to a Compute Engine VM behind Cloudflare. The container came up fine. DNS and a 521 error did not. Two failures, both preventable, both worth knowing before your first deploy.

We wired bidirectional sync into self-hosted Twenty CRM. Our own writes kept echoing back as inbound changes. The fix wasn't a clever filter — it was storing our API key identity once and comparing against it on every webhook.