Month 5 Plan

Month 5 — Security and Observability: Production Hardness

This month the system stops being a "trial project" and becomes a "deployable-to-production platform". An AI infrastructure engineer who knows both security and observability is rare and expensive.

Learning core (both of you)

• Service security: authentication vs authorization
• Token-based authentication
• Rate limiting
• AI-specific security: prompt injection, tool permissions
• Structured logging
• Metrics, monitoring, distributed tracing
• CI/CD

Concepts you must be able to explain

• Auth vs authz · [ ] Signed access token · [ ] Rate limiting algorithm
• Prompt injection · [ ] Tool permission · [ ] Structured log
• Counter/gauge/histogram · [ ] Trace and span · [ ] Correlation ID · [ ] CI/CD pipeline

Project steps

1. Authentication in control-plane
2. Authorization: a user sees only their own tasks
3. Rate limiting
4. Tool permissions + basic prompt-injection defense
5. Structured logging + correlation ID (carried across services)
6. Metrics dashboard (task count, success rate, latency distribution, token cost, error rate)
7. Distributed tracing: follow a task end to end, down to each step
8. CI/CD: tests + eval on every push
9. One-command setup (including the monitoring dashboard)
10. Chaos scenario: deliberately stop a service, document the behavior

Why is one-command setup a separate step? Because a person examining the repo takes the project seriously if they can run it. A project that requires pages of setup instructions gets closed without a try, no matter how good it is.

Division of labor

Split and swap at the rotation — you decide who takes what. Two rough tracks: (a) auth + authz + rate limit + metrics dashboard + chaos scenario, (b) structured logging + tracing + tool permissions + prompt-injection defense.

Content plan

• Observability dashboard tour
• Prompt injection demo ("how AI agents are fooled — and how they're protected")
• Chaos test clip
• CI/CD showcase post

Retro