What makes an MCP agent production-grade?
The difficult part is not giving a model tools. It is designing the permissions, evidence, evaluation, and operating model around those tools.
Writing
Engineering systems after the demo.
Notes on AI systems, automation, observability, and reliability—with an emphasis on what happens when the system meets real permissions, failures, ownership, and operating costs.
Subscribe via RSS →Automating engineering toil with n8n—without creating new toil
Workflow tools create leverage when the process, failure behavior, ownership, and role of AI are designed before the canvas fills with nodes.
Adding observability to an infrastructure stack
Start with the decisions operators need to make, then design signals, ownership, and cost controls that remain useful under stress.