Split monolithic AI automations into an orchestrator and reusable workers. Typed n8n sub-workflow contracts, Claude Code subagents, and the check most builds miss.
I build and execute my own workflows from Claude Cowork and ChatGPT Work using similar principles that you describe; my workflow execution md file is the orchestrator (with a bit more duties like global context loading) and each step in the workflow is typically a md file (sub-agent pattern).
The difference I guess is that I have the full session context (global) which doesn’t seem to be available through sub-agents.
I ran into validation logic issues too. It’s freaking hard in a non-deterministic system but also fun to figure out 😉
Agree that building is the easy part. I spend all my time testing and tweaking before pushing it to my students.
Very interesting!
I build and execute my own workflows from Claude Cowork and ChatGPT Work using similar principles that you describe; my workflow execution md file is the orchestrator (with a bit more duties like global context loading) and each step in the workflow is typically a md file (sub-agent pattern).
The difference I guess is that I have the full session context (global) which doesn’t seem to be available through sub-agents.
I ran into validation logic issues too. It’s freaking hard in a non-deterministic system but also fun to figure out 😉
Agree that building is the easy part. I spend all my time testing and tweaking before pushing it to my students.
Absolutely @Fabrice Talbot so glad to see the patterns and principles resonated and yeah I agree that is the fun part :D