Sticky-note appendix

Compressed for the wall above your monitor. The seven things we wish we had known on day one:

1. The agent runs as a different user. Permission-set every Apex class.
2. Stubs that return 'stub_response' poison the planner. Return real data or ''/'false'.
3. enabled_tools in the trace is the truth about what the LLM sees.
4. @outputs.X is what the planner says Apex returned. Defend against it lying with descriptions, flag-based gates, and correlated checks.
5. The reserved Apex word in will silently break your loop. Use inp.
6. if @variables.customer_id != "" is a dangerously loose gate. Any non-empty hallucination passes it. Gate on explicit flags from Apex.
7. --use-live-actions is mandatory when iterating against real backing logic. Without it, you are testing the LLM, not the system.