Liam McGarrigle is a senior developer advocate at n8n who designs practical safeguards for AI agents operating inside real business systems. His work combines human-in-the-loop automation, visual workflow development, scoped permissions, and tools that make agent behavior easier to inspect.
A self-taught developer, McGarrigle worked independently as an integration and automation consultant before joining n8n in his first full-time role. He had already contributed to its technical community, where he became a community moderator; his initial responsibilities at the company included education and certification.
His projects chart that progression from integration consulting to agent infrastructure. In 2024, he introduced n8n-auto-install, an unofficial self-hosting helper whose repository now warns that it needs security updates. He built community integrations for Zoho Calendar and Zoho Bookings, then explored workflow testing and evaluation through n8n-with-tests, a fork subsequently archived as official testing capabilities emerged. His newer agent-config-sync synchronizes skills and MCP configuration across Claude Code, OpenCode, Codex, and Cursor.
Building agents people can actually control
- Enforce approval at the tool boundary. His AI Engineer Europe workshop demonstrated Gmail and Google Calendar agents that cannot send messages or create events without explicit human authorization. Approval screens display proposed recipients, messages, and event details before execution.
- Restrict authority and preserve visibility. McGarrigle limits which tool fields an agent can populate, separates project credentials, exposes MCP workflows individually, and uses execution logs to investigate decisions and failures. Specialized subagents can divide responsibilities across different models without overloading a single context window.
- Make agent instructions modular. His agent-skills workflow indexes reusable skills from GitHub and retrieves relevant instructions when needed. His JSON-validation template checks incoming payloads and returns actionable errors without invoking AI, extending the same operational discipline beyond the model itself.
