AI
n8n as an Agentic Workflow Platform
n8n launched in 2019 as a low-code integration tool, years before AI agents became a common conversation. Triggers fire a workflow, actions talk to external apps like Gmail or Salesforce, and control-flow nodes branch on conditions. That foundation already worked before anyone attached an LLM to it.
The problem McGarrigle names is not building agents. It is knowing what your agent did, spotting what went wrong, and fixing it. n8n’s visual canvas exposes every step, every input, and every output. That observability is what makes it useful for agentic work rather than just another chat wrapper.
Wiring Up an AI Agent Node
A chat trigger feeds messages into an AI agent node, which needs a chat model and optionally memory. McGarrigle uses OpenRouter so attendees can pick any model, GPT-4o, Claude Sonnet 4.6, or others, without swapping credentials. Without a memory node, the agent forgets every prior turn. Connecting simple memory and raising the context window from the default 5 to 50 keeps the conversation coherent.
The built-in ChatHub interface appears as a sidebar icon once the workflow is published, giving a usable chat UI with no Slack or external tool required.
Tool Setup and Tool-Level Prompting
Each Gmail or Calendar tool field can be individually locked. Lock the “from” address and the agent can never change the sender. The trade-off is that every field the agent needs must be explicitly configured. Node names and descriptions go to the LLM verbatim as tool names and instructions, so McGarrigle writes tone and formatting rules in the node description rather than the system prompt, keeping each tool self-contained.
One concrete fix: Google Calendar’s API names the event title “summary,” so models write prose summaries. Renaming the field key to “title” corrects it.
Human-in-the-Loop Gates
The human review node intercepts tool calls before any API call goes out. The agent has no visibility into the gate. To prevent smarter models from hedging and refusing to call the tool, McGarrigle adds a line to the tool description:
“this won’t send automatically. Don’t be afraid to send a message. Message there is a human in the loop.” — Liam McGarrigle
He walks through the live send-email human review demo (42:44) and the live calendar event human review demo (54:11), fixing the UTC timestamp with a Luxon format string so reviewers see a readable date before approving.
Scheduling, Slack, and the Native MCP Server
Swapping the chat trigger for a schedule trigger set to once per hour turns the same workflow into an autonomous inbox manager. It still hits the human review gates, but requires no manual start. McGarrigle runs this for GitHub PRs: the agent scans new issues, drafts follow-up questions, and queues them for his approval before any message goes out.
n8n’s native MCP server (cloud and enterprise) exposes published workflows as callable tools. An agent like Claude Code connects via OAuth and can send chat messages to the running bot or execute any enabled workflow from the CLI.
Q&A
What is the new native n8n MCP server and how does it differ from the unofficial community one? The native server is built into n8n settings, connects via OAuth or access token, and can execute workflows (send a chat message to a running agent) in addition to creating and reading them, which the community MCP cannot do. ▶ 1:00:05
Do waiting executions (paused at a human review step) count against the concurrent execution limit on self-hosted instances? McGarrigle did not know for certain but suggested setting a timeout on the review node (for example, auto-deny after 10 minutes) to prevent waiting executions from piling up indefinitely. ▶ 1:08:28
Is there a PR-style approval flow for teams editing the same workflow? On non-enterprise plans, the workaround is copying the workflow, making changes, then manually replacing it. Enterprise plans have full git integration with dev, staging, and production environments and named approvers. ▶ 1:10:09
Can the human review gate work in headless contexts like phone calls where there is no button UI? The review node is a hard stop tied to chat platforms only. McGarrigle’s workaround is a subworkflow that handles the confirmation logic manually with an if-node, then returns true or false to the calling agent. ▶ 1:13:28
Notable Quotes
this won’t send automatically. Liam McGarrigle · ▶ 51:35
Don’t be afraid to send a message. Liam McGarrigle · ▶ 51:39
Message there is a human in the loop. Liam McGarrigle · ▶ 51:45
Key Takeaways
- Each Gmail or Calendar tool field can be individually locked or opened to AI, limiting what the agent can change.
- The human review node intercepts tool calls before execution. The agent cannot see or bypass it.
- A schedule trigger plus Slack replaces the chat trigger, turning a reactive agent into an autonomous hourly inbox manager.