APPENDIX G: LAB SETUP GUIDE

THE FIVE STAGES AT A GLANCE Stage 1 in Chapter 65: build the core loop, a tool registry with three tools, and a permission gate, using five fictional resumes and one fictional job posting. The agent must list the resumes, read each, compare it with the posting, and write a one paragraph report for each in the reports folder, with no more than ten steps per resume and no reading or writing outside the permitted folders. Stage 2 in Chapter 66: add the context builder, compaction, and the progress notes file, and grow the sample data to twenty-five resumes, including at least two unusually long ones. The full batch must stay within a token budget, and after a pretend restart the agent must continue from its notes without redoing finished resumes. Stage 3 in Chapter 67: connect to a local Model Context Protocol server, add a human approval step before an existing report can be overwritten, and defend against the planted injection. You must find the hidden instruction in resume_injection_test.txt and the missing idempotency check in write_report, and you may try the optional stretch bug of a poisoned tool description. Stage 4 in Chapter 68: add tracing and a dashboard, and build the fifteen item evaluation set. You must find the runaway loop triggered by resume_empty.txt, using the repeated pattern in the trace, and fix it with a guard against repeated identical calls. Stage 5 in Chapter 69: split the agent into a supervisor with a matcher and a writer, route the matcher to a cheaper model and the writer to a stronger one, add retries with exponential backoff and a maximum iteration limit, and fill in an honest production-readiness checklist.